Research collection on agent memory architectures, persistence patterns, and output quality maintenance for LLM-based agent systems.
If you reference this repo’s summaries/analyses in academic or professional work, please cite:
@misc{lin_agentic_memory_2026,
author = {Leonard Lin},
title = {agentic-memory: Agentic Memory Research Collection (Summaries and Analyses)},
year = {2026},
howpublished = {GitHub repository},
url = {https://github.com/lhl/agentic-memory},
}| Document | Author | Description |
|---|---|---|
| jumperz-agent-memory-stack | @jumperz | 31-piece memory architecture split across 3 phases (Core → Reliability → Intelligence). Complete prompt/spec breakdowns for write pipeline, read pipeline, decay, knowledge graph, episodic memory, trust scoring, echo/fizzle feedback loops. The foundational reference that others build on. |
| joelhooks-adr-0077-memory-system-next-phase | @joelhooks | ADR for joelclaw (personal AI Mac Mini). Maps existing production system (~6 days running, Qdrant 1,343 points) against jumperz's 31 pieces. Plans 3 increments: retrieval quality (score decay, query rewriting), storage quality (dedup, nightly maintenance), feedback loop (echo/fizzle). Includes detailed gap analysis. |
| coolmanns-openclaw-memory-architecture | coolmanns | 12-layer production memory stack for OpenClaw with 14 agents. SQLite+FTS5 knowledge graph (3,108 facts), llama.cpp GPU embeddings (768d, 7ms), three runtime plugins (continuity, stability, graph-memory). 100% recall on 60-query benchmark. Includes activation/decay system, domain RAG, session boot sequences. |
| drag88-agent-output-degradation | @drag88 (Aswin) | "Why Your Agent's Output Gets Worse Over Time" — multi-agent convergence problem. 4-tier memory (working → episodic → semantic → procedural). 3-layer enforcement pipeline (YAML regex → Gemini LLM judge → self-learning loop). Core insight: convert expensive runtime LLM checks into free static regex rules over time. |
| versatly-clawvault | Versatly (@drag88) | ClawVault npm CLI tool — structured markdown memory vault with observation pipeline, knowledge graph, session lifecycle (wake/sleep/checkpoint), task/project primitives, Obsidian integration, OpenClaw hooks. 449+ tests. v2.6.1. |
| vstorm-memv | vstorm-co | memv (PyPI: memvee) — Nemori-inspired predict-calibrate extraction + episode segmentation, plus Graphiti-style bi-temporal validity and hybrid retrieval (sqlite-vec + FTS5 + RRF) on SQLite. |
| supermemory | Dhravya Shah / supermemoryai | Supermemory memory-as-a-service API: memory versioning (linked-list chains), typed relationships (updates/extends/derives), static/dynamic profile synthesis, time-based forgetting with reason tracking, multi-model embedding storage. Critical caveat: open-source repo is frontend/SDK only; core engine is proprietary backend at api.supermemory.ai. |
| Document | Author | Description |
|---|---|---|
| hu-evermembench | Hu et al. | EverMemBench benchmark for >1M-token multi-party, multi-group interleaved conversations; diagnoses multi-hop collapse, temporal/versioning difficulty, and retrieval-bottlenecked “memory awareness”. |
| zhang-live-evo | Zhang et al. | Live-Evo: online self-evolving agent memory with an experience bank + meta-guideline bank, contrastive “memory-on vs memory-off” feedback, and weight-based reinforcement/forgetting; evaluated on Prophet Arena + deep research (as reported). |
| shutova-structmemeval | Shutova et al. | StructMemEval benchmark for whether agents can organize memory into useful structures (trees/ledgers/state tracking), not just retrieve facts; includes hint vs no-hint evaluation to isolate “structure recognition” failures. |
| yan-gam | Yan et al. | GAM: just-in-time agent memory via lightweight memos + a universal page-store, plus a deep-research researcher that plans/searches/integrates/reflects over history to compile optimized context at runtime; strong long-context QA gains with higher latency (as reported). |
| yang-graph-based-agent-memory-taxonomy | Yang et al. | Graph-based Agent Memory survey: graph-centric taxonomy + lifecycle (extract/store/retrieve/evolve), storage structures (KG/temporal/hyper/hierarchical/hybrid), retrieval operators, evolution/maintenance, and resources/benchmarks; useful shared vocabulary for shisad. |
| zhang-survey-memory-mechanism | Zhang et al. | Survey on memory mechanisms for LLM agents: definitions, why memory, design axes (sources/forms/ops), evaluation approaches, and application domains; good baseline checklist alongside newer benchmarks/systems. |
| hu-memory-age-ai-agents | Hu et al. | Memory in the Age of AI Agents survey: proposes unified lenses of forms (token/parametric/latent), functions (factual/experiential/working), and dynamics (formation/evolution/retrieval), plus benchmarks/frameworks and trustworthiness frontiers. |
| li-locomoplus | Li et al. | LoCoMo-Plus: evaluates beyond-factual “cognitive memory” (latent constraints like state/goals/values) under cue–trigger semantic disconnect, using constraint-consistency + LLM-judge evaluation. |
| maharana-locomo | Maharana et al. | LoCoMo dataset + benchmark for very long-term multi-session conversations (300 turns, multimodal) grounded in personas + temporal event graphs; evaluates QA + event summarization + multimodal generation. |
| wu-longmemeval | Wu et al. | LongMemEval benchmark + design decomposition (indexing → retrieval → reading) and system optimizations (value granularity, key expansion, time-aware query expansion). |
| packer-memgpt | Packer et al. | MemGPT: OS-inspired hierarchical memory + paging between a fixed-context LLM prompt and external stores (recall + archival), with function-call memory ops and event-driven control flow; foundational baseline for external agent memory. |
| chhikara-mem0 | Chhikara et al. | Mem0: production-oriented long-term memory pipeline with explicit ops (ADD/UPDATE/DELETE/NOOP) and an optional graph memory variant; reports quality + token/latency tradeoffs on LoCoMo. |
| liu-simplemem | Liu et al. | SimpleMem: write-time semantic structured compression + online synthesis + intent-aware retrieval planning (multi-view dense/BM25/symbolic retrieval with union+dedup) to improve LoCoMo/LongMemEval quality while cutting token cost (as reported). |
| xu-a-mem | Xu et al. | A‑Mem: Zettelkasten-inspired note network with LLM-driven link generation and “memory evolution” (updating older note attributes as new evidence arrives); strong LoCoMo multi-hop/temporal gains with far lower token lengths than full-context (as reported). |
| salama-meminsight | Salama et al. | MemInsight: autonomous memory augmentation that mines/annotates attributes (entity-centric + conversation-centric; turn/session granularity) and uses attribute-guided retrieval; large LoCoMo retrieval recall gains vs DPR RAG baseline (as reported). |
| rasmussen-zep | Rasmussen et al. | Zep: production memory layer built on Graphiti, a bi-temporal knowledge graph (episodes → entities/facts → communities) with validity intervals and invalidation-based corrections; evaluated on DMR + LongMemEval. |
| nan-nemori | Nan et al. | Nemori: cognitively-inspired self-organizing agent memory with semantic episode boundary detection + episodic narratives and a predict-calibrate loop that distills semantic knowledge from prediction gaps; strong LoCoMo + LongMemEvalS results (as reported). |
| li-memos | Li et al. | MemOS: OS-like memory control plane with MemCube (payload+metadata), lifecycle/scheduling, governance (ACL/TTL/audit), and multi-substrate memory (plaintext/activation/KV/parameter/LoRA). |
| yan-memory-r1 | Yan et al. | Memory-R1: reinforcement-learned memory manager (ADD/UPDATE/DELETE/NOOP) + answer agent with learned memory distillation; data-efficient RL (PPO/GRPO) training with exact-match reward. |
| jonelagadda-mnemosyne | Jonelagadda et al. | Mnemosyne: edge-friendly graph memory with substance/redundancy filters, probabilistic recall with decay/refresh, and a fixed-budget “core summary” for persona-level context. |
| patel-engram | Patel et al. | ENGRAM: lightweight typed memory (episodic/semantic/procedural) with simple dense retrieval + strict evidence budgets; strong LoCoMo + LongMemEval results with low token usage. |
| wei-evo-memory | Wei et al. | Evo-Memory: streaming benchmark + framework for self-evolving memory and experience reuse; introduces ExpRAG and ReMem (Think/Act/Refine) baselines and robustness/efficiency metrics. |
| cao-remember-me-refine-me | Cao et al. | ReMe: dynamic procedural memory lifecycle (acquire→reuse→refine) with multi-faceted distillation from success/failure trajectories, scenario-aware retrieval, and utility-based pruning; strong BFCL‑V3/AppWorld results (as reported). |
| sarin-memoria | Sarin et al. | Memoria: personalization memory layer combining session summaries + KG triplets (persona) with exponential recency weighting; SQLite + ChromaDB architecture and LongMemEvals subset results. |
| latimer-hindsight | Latimer et al. | Hindsight: retain/recall/reflect architecture separating evidence vs beliefs vs summaries; temporal+entity memory graph with multi-channel retrieval fusion and belief confidence updates; very strong LongMemEval/LoCoMo results (as reported). |
| yu-agentic-memory | Yu et al. | AgeMem: RL-trained unified LTM+STM controller exposing memory ops as tool actions (add/update/delete/retrieve/summarize/filter) with a 3-stage curriculum and step-wise GRPO for credit assignment. |
| hu-evermemos | Hu et al. | EverMemOS: self-organizing “memory OS” with MemCells→MemScenes lifecycle, user profile consolidation, and necessity/sufficiency-guided recollection (verifier + query rewrite); strong LoCoMo/LongMemEval results (as reported). |
| li-timem | Li et al. | TiMem: temporal-hierarchical memory consolidation (segment→session→day→week→profile) with query-complexity recall planning + gating; strong LoCoMo/LongMemEval-S accuracy with low recalled tokens (as reported). |
| zhang-himem | Zhang et al. | HiMem: hierarchical long-term memory split (Episode Memory + Note Memory) with topic+surprise episode segmentation, note-first “best-effort” retrieval w/ sufficiency checks, and conflict-aware reconsolidation; strong LoCoMo results (as reported). |
| behrouz-nested-learning | Behrouz et al. | Nested Learning / CMS / Hope: reframes memory as multi-timescale update dynamics (continuum memory blocks updated at different frequencies) with implications for consolidation and “corrections without forgetting”. |
| zhang-recursive-language-models | Zhang et al. | Recursive Language Models (RLMs): inference-time recursion + REPL state treats long prompts as an external environment; processes multi‑million-token inputs with sub-calls and programmatic slicing, often beating long-context scaffolds at comparable average cost (as reported). |
| wang-m-plus | Wang et al. | M+: latent-space long-term memory extension to MemoryLLM that stores dropped memory tokens in an LTM pool and retrieves them during generation with a co-trained retriever; extends retention to >160k tokens at similar GPU memory cost (as reported). |
| dong-minja | Dong et al. | MINJA: practical memory injection attack on “memory-as-demonstrations” agents via query-only interaction (bridging steps + progressive shortening); motivates write-time gates, isolation, and safer memory representations. |
| sunil-memory-poisoning-attack-defense | Sunil et al. | Memory poisoning attack & defense: empirical MINJA follow-up in EHR agents; shows pre-existing benign memory can reduce ASR, and that trust-score defenses can fail via over-conservatism or overconfidence. |
| anokhin-arigraph | Anokhin et al. | AriGraph: knowledge-graph world model that links episodic observation nodes to extracted semantic triplets; two-stage retrieval (semantic→episodic) for planning/exploration in text-game environments. |
| behrouz-titans | Behrouz et al. | Titans: long-context architecture with an online-updated neural memory module (test-time learning) plus persistent task memory; provides explicit primitives for surprise-based salience and forgetting. |
| ahn-hema | Ahn | HEMA: hippocampus-inspired dual memory for long conversations (running compact summary + FAISS episodic vector store) with explicit prompt budgeting, pruning (“semantic forgetting”), and summary-of-summaries consolidation. |
| tan-membench | Tan et al. | MemBench: benchmark/dataset for agent memory covering participation vs observation scenarios and factual vs reflective memory, with metrics for accuracy/recall/capacity and read/write-time efficiency. |
Root-level critical analyses intended for synthesis work. These reference the summaries above, but focus on coherence, evidence quality, risks, and synthesis-ready claim framing.
| Synthesis | Based on | Focus |
|---|---|---|
| ANALYSIS | ANALYSIS-*.md + shisad docs + Mem0/Letta baselines |
Cross-system comparison (techniques + memory types), plus mapping to shisad and “traditional” RAG-ish memory |
| ANALYSIS-academic-industry | paper ANALYSIS-arxiv-*.md + shisad plan |
Academic/industry synthesis: benchmarks vs systems vs attacks, with “what’s missing in shisad” framing |
| Benchmarks best practices | Public disputes, audits, our analysis | Known pitfalls, metric confusion, dataset quality issues, per-benchmark limitations |
| MELT benchmark design | ANALYSIS.md systems + Reality Check epistemic docs | Memory Evaluation for Lifecycle Testing — session-replay benchmark testing full memory lifecycle (decay, consolidation, contradiction, core stability, inference) at 6 scale tiers over simulated time. Separate repo; draft. |
| Analysis | Based on | Focus |
|---|---|---|
| ANALYSIS-jumperz-agent-memory-stack | references/jumperz-agent-memory-stack.md |
Checklist critique (semantics, failure modes, missing evaluation), synthesis-ready takeaways + claims table |
| ANALYSIS-joelhooks-adr-0077-memory-system-next-phase | references/joelhooks-adr-0077-memory-system-next-phase.md |
Increment plan critique (decay, rewrite, dedup, echo/fizzle), validation plan + claims |
| ANALYSIS-coolmanns-openclaw-memory-architecture | references/coolmanns-openclaw-memory-architecture.md + vendor/openclaw-memory-architecture/ |
Layered stack critique with benchmark-method verification, operational risks, doc drift notes |
| ANALYSIS-drag88-agent-output-degradation | references/drag88-agent-output-degradation.md |
Convergence + enforcement pattern critique (judge→rule distillation), measurement gaps, risks |
| ANALYSIS-versatly-clawvault | references/versatly-clawvault.md + vendor/clawvault/ |
Product/tooling critique (surface area, hooks, qmd dependency), security posture, missing benchmarks |
| ANALYSIS-vstorm-memv | references/vstorm-memv.md + vendor/memv/ |
Implementation critique of Nemori-inspired predict-calibrate extraction + bi-temporal validity + hybrid retrieval, with gaps/risks and shisad mapping |
| ANALYSIS-openviking | vendor/openviking/ + Hermes provider docs |
Open-source context database: viking:// filesystem, L0/L1/L2 tiered loading, session-commit extraction across 8 memory categories, and hierarchical typed retrieval over memory/resources/skills; strong observability with heavier operational complexity |
| ANALYSIS-byterover-cli | vendor/byterover-cli/ + vendor/byterover-cli/paper/ |
Agent-native coding-agent memory/runtime: daemon + per-project agent pool, markdown context tree with explicit relations and lifecycle, 5-tier progressive retrieval with cache/OOD detection, and strong self-reported benchmarks with caveats |
| ANALYSIS-mira-OSS | vendor/mira-OSS/ |
Full-stack event-driven agent (v1 rev 2): activity-day sigmoid decay, hub discovery + 3-axis linking (vector+entity+TF-IDF), Text-Based LoRA + user model synthesis with critic validation, background forage agent (sub-agent collaboration), portrait synthesis, 16 tools, context overflow remediation, immutable domain models, multi-user RLS + Vault; gaps in write gating, external benchmarks, taint tracking, and sub-agent capability scoping |
| ANALYSIS-claude-code-memory | Source: /home/lhl/Downloads/claude-code/src |
Claude Code memory subsystem (Anthropic): first-party production-scale memory system; flat-file MEMORY.md + typed topic files (user/feedback/project/reference) + background extraction via forked agent with mutual exclusion + LLM-based relevance selection (Sonnet) + team memory with OAuth sync + auto dream consolidation + KAIROS daily-log mode + eval-validated prompts with case IDs + security-hardened path validation; no vector search, no graph, no decay scoring |
| ANALYSIS-codex-memory | openai/codex | Codex memory subsystem (OpenAI): first-party open-source coding agent; two-phase async pipeline (gpt-5.1-codex-mini extraction → gpt-5.3-codex consolidation) + SQLite-backed job coordination (leases/heartbeats/watermarks) + progressive disclosure layout (memory_summary → MEMORY.md → rollout_summaries → skills) + skills as procedural memory + usage-based citation-driven retention + thread-diff incremental forgetting + ~1,400 lines extraction/consolidation prompts; no vector search, no team memory, no real-time extraction |
| ANALYSIS-google-always-on-memory-agent | vendor/always-on-memory-agent/ |
Official Google ADK sample: always-on daemon with multimodal ingestion (27 file types via Gemini 3.1 Flash-Lite), periodic LLM consolidation, SQLite storage, HTTP API + Streamlit dashboard; no retrieval/search (recency scan LIMIT 50), no decay/dedup/versioning; useful as ADK orchestration reference and multimodal ingestion pattern |
| ANALYSIS-supermemory | references/supermemory.md + vendor/supermemory/ |
Memory-as-a-service startup: memory versioning (linked-list chains via parentMemoryId/rootMemoryId/isLatest), typed relationship ontology (updates/extends/derives), static/dynamic profile synthesis API, time-based forgetting with audit trail, multi-model embedding columns, MemoryBench framework; open-source repo is SDK/frontend only — core engine logic is proprietary hosted backend |
| ANALYSIS-karta | vendor/karta/ |
Karta (rohithzr): Rust (~10.4K LOC) agentic memory library with Zettelkasten-inspired knowledge graph, 7-type dream engine (deduction/induction/abduction/consolidation/contradiction/episode digest/cross-episode digest) with inference feedback into retrieval, embedding-based query classification (6 modes), retroactive context evolution with drift protection, cross-encoder reranking with abstention, multi-hop BFS traversal, atomic fact decomposition with per-fact embeddings, foresight signals with TTL, structured episode digests; BEAM 100K: 57.7% with 243-failure root cause catalog |
| Analysis | Based on | Focus |
|---|---|---|
| ANALYSIS-arxiv-2602.01313-evermembench | references/hu-evermembench.md + references/papers/arxiv-2602.01313.pdf |
Benchmark critique emphasizing version semantics, multi-party fragmentation, oracle diagnostics, and shisad mapping |
| ANALYSIS-arxiv-2602.02369-live-evo | references/zhang-live-evo.md + references/papers/arxiv-2602.02369.pdf |
System deep dive emphasizing online experience weighting from continuous feedback, meta-guidelines for memory compilation, and memory-on vs memory-off utility measurement; shisad mapping for feedback loops + procedural memory gating |
| ANALYSIS-arxiv-2602.11243-structmemeval | references/shutova-structmemeval.md + references/papers/arxiv-2602.11243.pdf |
Benchmark deep dive emphasizing memory organization/structure as a distinct capability (trees/ledgers/state), hint vs no-hint diagnostics, and implications for shisad structured-memory primitives |
| ANALYSIS-arxiv-2602.05665-graph-based-agent-memory-taxonomy | references/yang-graph-based-agent-memory-taxonomy.md + references/papers/arxiv-2602.05665.pdf |
Survey deep dive providing graph-based memory taxonomy and lifecycle (extract/store/retrieve/evolve), with implications for shisad graph-as-derived-view, operator choices, and maintenance jobs |
| ANALYSIS-arxiv-2404.13501-survey-memory-mechanism | references/zhang-survey-memory-mechanism.md + references/papers/arxiv-2404.13501.pdf |
Survey deep dive providing baseline taxonomy and evaluation checklists for agent memory; useful coverage reference alongside newer benchmarks/systems for shisad’s roadmap |
| ANALYSIS-arxiv-2512.13564-memory-age-ai-agents | references/hu-memory-age-ai-agents.md + references/papers/arxiv-2512.13564.pdf |
Survey deep dive emphasizing the Forms–Functions–Dynamics taxonomy and frontiers (RL integration, multimodal, multi-agent shared memory, trustworthiness), used as organizing frame for shisad v0.7 memory roadmap |
| ANALYSIS-arxiv-2402.17753-locomo | references/maharana-locomo.md + references/papers/arxiv-2402.17753.pdf |
Dataset/benchmark critique with episodic-memory implications (event graphs, multimodal, RAG harm) and shisad mapping |
| ANALYSIS-arxiv-2410.10813-longmemeval | references/wu-longmemeval.md + references/papers/arxiv-2410.10813.pdf |
Benchmark and system-design decomposition (indexing/retrieval/reading), with mapping to shisad primitives |
| ANALYSIS-arxiv-2310.08560-memgpt | references/packer-memgpt.md + references/papers/arxiv-2310.08560.pdf |
System deep dive emphasizing virtual context management (OS paging), memory tiers (working/queue/recall/archival), function-call memory ops, and implications for shisad versioned corrections + write-policy hardening |
| ANALYSIS-arxiv-2602.10715-locomoplus | references/li-locomoplus.md + references/papers/arxiv-2602.10715.pdf |
Beyond-factual “cognitive memory” benchmark critique (latent constraints) and implications for safe constraint/procedural memory |
| ANALYSIS-arxiv-2504.19413-mem0 | references/chhikara-mem0.md + references/papers/arxiv-2504.19413.pdf |
System deep dive emphasizing explicit memory ops, graph-memory tradeoffs, deployment metrics (tokens/p95), and shisad mapping (versioned corrections vs delete) |
| ANALYSIS-arxiv-2601.02553-simplemem | references/liu-simplemem.md + references/papers/arxiv-2601.02553.pdf |
System deep dive emphasizing write-time semantic structured compression, online consolidation, and intent-aware multi-view retrieval planning; mapping to shisad “derived vs raw” memory + retrieval budgeting |
| ANALYSIS-arxiv-2502.12110-a-mem | references/xu-a-mem.md + references/papers/arxiv-2502.12110.pdf |
System deep dive emphasizing Zettelkasten-style notes + LLM-driven linking + memory evolution, with strong multi-hop/temporal LoCoMo gains but high versioning/audit requirements for shisad |
| ANALYSIS-arxiv-2503.21760-meminsight | references/salama-meminsight.md + references/papers/arxiv-2503.21760.pdf |
System deep dive emphasizing autonomous attribute mining/annotation as a derived metadata layer to improve retrieval recall and downstream tasks; mapping to shisad schema constraints + provenance/versioning |
| ANALYSIS-arxiv-2511.18423-gam | references/yan-gam.md + references/papers/arxiv-2511.18423.pdf |
System deep dive emphasizing just-in-time context compilation via memo index + universal page-store and an iterative deep-research researcher; highlights the latency/quality trade-off and mapping to shisad evidence-first episodic storage |
| ANALYSIS-arxiv-2501.13956-zep | references/rasmussen-zep.md + references/papers/arxiv-2501.13956.pdf |
System deep dive emphasizing bi-temporal validity semantics, episodic+semantic+community graph tiers, hybrid retrieval (BM25/embeddings/BFS), and implications for shisad versioned memory |
| ANALYSIS-arxiv-2507.03724-memos | references/li-memos.md + references/papers/arxiv-2507.03724.pdf |
System deep dive emphasizing MemCube metadata, multi-substrate memory (plaintext/KV/parameter), lifecycle/scheduling/governance, and mapping to shisad primitives |
| ANALYSIS-arxiv-2508.19828-memory-r1 | references/yan-memory-r1.md + references/papers/arxiv-2508.19828.pdf |
RL deep dive emphasizing learned memory ops (ADD/UPDATE/DELETE/NOOP) + post-retrieval memory distillation, reward design, and what’s required to safely adopt this in shisad |
| ANALYSIS-arxiv-2508.03341-nemori | references/nan-nemori.md + references/papers/arxiv-2508.03341.pdf |
System deep dive emphasizing episode segmentation (Two-Step Alignment) + predict-calibrate semantic distillation, reported LoCoMo/LongMemEvalS gains, and implications for shisad write gating + correction semantics |
| ANALYSIS-arxiv-2510.08601-mnemosyne | references/jonelagadda-mnemosyne.md + references/papers/arxiv-2510.08601.pdf |
System deep dive emphasizing edge-first graph memory, redundancy/refresh, probabilistic decay-based recall, and a fixed-budget core/persona summary; includes evaluation-rigor cautions |
| ANALYSIS-arxiv-2511.12960-engram | references/patel-engram.md + references/papers/arxiv-2511.12960.pdf |
System deep dive emphasizing typed memory (episodic/semantic/procedural), deterministic routing/formatting, strict evidence budgets, and strong token/latency results; mapping to shisad primitives |
| ANALYSIS-arxiv-2511.20857-evo-memory | references/wei-evo-memory.md + references/papers/arxiv-2511.20857.pdf |
Benchmark deep dive emphasizing streaming task-sequence evaluation for experience reuse, plus refine/prune mechanisms and metrics (robustness, step efficiency) for shisad’s eval harness |
| ANALYSIS-arxiv-2512.10696-remember-me-refine-me | references/cao-remember-me-refine-me.md + references/papers/arxiv-2512.10696.pdf |
System deep dive emphasizing procedural memory distillation + scenario-aware reuse + utility-based refinement/pruning; mapping to shisad procedural tier + versioned invalidation vs delete |
| ANALYSIS-arxiv-2512.12686-memoria | references/sarin-memoria.md + references/papers/arxiv-2512.12686.pdf |
System deep dive emphasizing persona KG + session summaries with recency-weighted retrieval; highlights missing governance/versioning primitives needed for shisad |
| ANALYSIS-arxiv-2512.12818-hindsight | references/latimer-hindsight.md + references/papers/arxiv-2512.12818.pdf |
System deep dive emphasizing retain/recall/reflect with four-network memory (facts/experiences/observations/beliefs), token-budgeted multi-channel retrieval fusion, and belief confidence updates; key shisad mapping |
| ANALYSIS-arxiv-2601.01885-agentic-memory | references/yu-agentic-memory.md + references/papers/arxiv-2601.01885.pdf |
RL deep dive emphasizing unified LTM+STM memory ops as tool actions, 3-stage training curriculum, step-wise GRPO credit assignment, and implications for shisad’s future learned memory policies |
| ANALYSIS-arxiv-2601.02163-evermemos | references/hu-evermemos.md + references/papers/arxiv-2601.02163.pdf |
System deep dive emphasizing MemCell→MemScene consolidation lifecycle, user profile/foresight, and sufficiency-verified scene-guided retrieval; mapping to shisad consolidation roadmap |
| ANALYSIS-arxiv-2601.02845-timem | references/li-timem.md + references/papers/arxiv-2601.02845.pdf |
System deep dive emphasizing temporal-hierarchical consolidation (TMT), query-complexity recall planning/gating, and the accuracy–token frontier; mapping to shisad temporal tiers |
| ANALYSIS-arxiv-2601.06377-himem | references/zhang-himem.md + references/papers/arxiv-2601.06377.pdf |
System deep dive emphasizing Episode Memory + Note Memory hierarchy, note-first “best-effort” retrieval w/ sufficiency checks, and conflict-aware reconsolidation; mapping to shisad event→knowledge tiers + versioned updates |
| ANALYSIS-arxiv-2512.24695-nested-learning | references/behrouz-nested-learning.md + references/papers/arxiv-2512.24695.pdf |
Conceptual deep dive on multi-timescale “continuum memory” and consolidation dynamics; mapping to shisad tiered memory + versioned corrections |
| ANALYSIS-arxiv-2512.24601-recursive-language-models | references/zhang-recursive-language-models.md + references/papers/arxiv-2512.24601.pdf |
Architecture deep dive emphasizing RLM-style programmatic reading/compilation over arbitrarily long evidence stores (REPL + recursion + sub-calls), with implications for shisad sandboxed compilation traces and cost tail management |
| ANALYSIS-arxiv-2502.00592-m-plus | references/wang-m-plus.md + references/papers/arxiv-2502.00592.pdf |
Architecture deep dive emphasizing latent-space long-term memory tokens + co-trained retrieval for >160k retention, with mapping to shisad’s external evidence-first memory and retrieval diagnostics |
| ANALYSIS-arxiv-2503.03704-minja | references/dong-minja.md + references/papers/arxiv-2503.03704.pdf |
Security deep dive on query-only memory injection attacks; implications for write-policy, provenance/taint, isolation, and “don’t store demonstrations” patterns |
| ANALYSIS-arxiv-2601.05504-memory-poisoning-attack-defense | references/sunil-memory-poisoning-attack-defense.md + references/papers/arxiv-2601.05504.pdf |
Security deep dive emphasizing ISR vs ASR under realistic memory conditions, and why trust-score sanitization can fail; concrete shisad hardening takeaways |
| ANALYSIS-arxiv-2407.04363-arigraph | references/anokhin-arigraph.md + references/papers/arxiv-2407.04363.pdf |
System deep dive emphasizing episodic↔semantic memory linking, graph-structured retrieval for planning/exploration, and implications for shisad episode objects + provenance + correction semantics |
| ANALYSIS-arxiv-2501.00663-titans | references/behrouz-titans.md + references/papers/arxiv-2501.00663.pdf |
Architecture deep dive emphasizing test-time-learning neural memory (surprise/momentum/forgetting), Titans MAC/MAG/MAL variants, and how to translate salience/decay ideas into shisad’s external memory framework |
| ANALYSIS-arxiv-2504.16754-hema | references/ahn-hema.md + references/papers/arxiv-2504.16754.pdf |
System deep dive emphasizing dual memory (summary + vector store), explicit prompt budgeting, pruning/consolidation policies, and evaluation-rigor cautions for shisad adoption |
| ANALYSIS-arxiv-2506.21605-membench | references/tan-membench.md + references/papers/arxiv-2506.21605.pdf |
Benchmark deep dive emphasizing multi-scenario (participant vs observer) and multi-level (factual vs reflective) evaluation, plus latency/capacity metrics and implications for shisad eval harnesses |
| Source | URL |
|---|---|
| @jumperz memory stack thread | https://x.com/jumperz/status/2024841165774717031 |
| @joelhooks ADR tweet | https://x.com/joelhooks/status/2024947701738262773 |
| joelclaw ADR-0077 | https://joelclaw.com/adrs/0077-memory-system-next-phase |
| @drag88 article | https://x.com/drag88/status/2022551759491862974 |
| supermemory docs | https://supermemory.ai/docs |
| supermemory repo | https://github.com/supermemoryai/supermemory |
| mempalace repo | https://github.com/milla-jovovich/mempalace |
| karta repo | https://github.com/rohithzr/karta |
agentic-memory/
├── README.md ← this file
├── ANALYSIS.md ← synthesis + comparison
├── ANALYSIS-academic-industry.md ← academic/industry synthesis
├── ANALYSIS-jumperz-agent-memory-stack.md
├── ANALYSIS-joelhooks-adr-0077-memory-system-next-phase.md
├── ANALYSIS-coolmanns-openclaw-memory-architecture.md
├── ANALYSIS-drag88-agent-output-degradation.md
├── ANALYSIS-versatly-clawvault.md
├── ANALYSIS-vstorm-memv.md
├── ANALYSIS-mira-OSS.md
├── ANALYSIS-codex-memory.md
├── ANALYSIS-google-always-on-memory-agent.md
├── ANALYSIS-supermemory.md
├── ANALYSIS-karta.md ← Karta: Rust agentic memory library with dream engine
├── ANALYSIS-mempalace.md ← not in ANALYSIS.md (claims-vs-code issues); see REVIEWED.md
├── REVIEWED.md ← triage log (examined but not promoted to ANALYSIS)
├── PUNCHLIST-academic-industry.md ← tracking checklist for paper deep dives
├── templates/ ← templates for paper analyses/summaries
│
├── references/ ← summarized reference docs (markdown w/ frontmatter)
│ ├── 1-full-agent-memory-build.jpg ← jumperz card 1: memory storage
│ ├── 2-feeds-into.jpg ← jumperz card 2: memory intelligence
│ ├── jumperz-agent-memory-stack.md
│ ├── joelhooks-adr-0077-memory-system-next-phase.md
│ ├── coolmanns-openclaw-memory-architecture.md
│ ├── drag88-agent-output-degradation.md
│ └── versatly-clawvault.md
│ ├── hu-evermembench.md
│ ├── li-locomoplus.md
│ ├── maharana-locomo.md
│ ├── wu-longmemeval.md
│ ├── chhikara-mem0.md
│ └── papers/ ← archived PDFs + text snapshots
│ ├── README.md
│ ├── arxiv-*.pdf
│ └── arxiv-*.md
│
└── vendor/ ← cloned source repos
├── mira-OSS/ ← github.com/taylorsatula/mira-OSS (snapshot, AGPLv3)
│ ├── README.md
│ ├── CLAUDE.md ← project guide (architecture, patterns, principles)
│ ├── main.py ← FastAPI entry point
│ ├── cns/ ← Central Nervous System (conversation orchestration)
│ │ ├── api/ ← FastAPI endpoints (chat, actions, data, health)
│ │ ├── core/ ← Domain models (Continuum, Message, Events)
│ │ ├── services/ ← Orchestrator, subcortical, summary, collapse handler
│ │ └── infrastructure/ ← Repositories, Valkey cache, unit of work
│ ├── lt_memory/ ← Long-term memory system
│ │ ├── scoring_formula.sql ← Multi-factor activity-day sigmoid importance scoring
│ │ ├── models.py ← Memory, Entity, ExtractedMemory, link types
│ │ ├── hybrid_search.py ← BM25 + pgvector with RRF
│ │ ├── proactive.py ← Dual-path retrieval (similarity + hub discovery)
│ │ ├── hub_discovery.py ← Entity-driven memory retrieval via pg_trgm
│ │ └── processing/ ← Extraction, consolidation, entity GC pipelines
│ ├── working_memory/ ← System prompt composition via trinkets
│ ├── tools/ ← Self-registering tool framework (11 built-in)
│ ├── config/ ← Pydantic config + prompt templates
│ └── auth/ ← WebAuthn + magic link authentication
│
├── openclaw-memory-architecture/ ← github.com/coolmanns/openclaw-memory-architecture
│ ├── README.md
│ ├── PROJECT.md
│ ├── CHANGELOG.md
│ ├── docs/
│ │ ├── ARCHITECTURE.md ← full 12-layer technical reference
│ │ ├── knowledge-graph.md ← graph search pipeline, benchmarks
│ │ ├── context-optimization.md
│ │ ├── embedding-setup.md
│ │ ├── benchmark-process.md
│ │ ├── benchmark-results.md
│ │ ├── code-search.md
│ │ └── COMPARISON.md
│ ├── schema/
│ │ └── facts.sql ← SQLite schema for knowledge graph
│ ├── scripts/ ← init, seed, search, ingest, decay, benchmark, telemetry
│ ├── templates/ ← starter files (active-context, gating-policies, etc.)
│ └── plugin-graph-memory/ ← OpenClaw plugin (JS)
│
├── karta/ ← github.com/rohithzr/karta (submodule, MIT)
│ ├── Cargo.toml ← workspace: karta-core + karta-cli
│ ├── crates/
│ │ └── karta-core/ ← Core engine (~6.7K LOC Rust)
│ │ ├── src/
│ │ │ ├── note.rs ← MemoryNote, Provenance, NoteStatus, AtomicFact, Episode, EpisodeDigest
│ │ │ ├── write.rs ← Write path: index, link, evolve, foresight, facts
│ │ │ ├── read.rs ← Read path: classify, search, traverse, rerank, synthesize
│ │ │ ├── rerank.rs ← Jina/LLM/noop rerankers
│ │ │ ├── dream/ ← Dream engine: 7 inference types
│ │ │ ├── store/ ← LanceDB + SQLite implementations
│ │ │ └── llm/ ← Provider trait + OpenAI + mock + prompts
│ │ └── tests/ ← eval, beam_100k, bench_beam (~3.8K LOC)
│ ├── findings.md ← BEAM 100K detailed failure analysis
│ └── plan.md ← Experiment plan targeting 90%+
│
├── always-on-memory-agent/ ← GoogleCloudPlatform/generative-ai (official ADK sample)
│ ├── agent.py ← ADK multi-agent daemon (ingest/consolidate/query)
│ ├── dashboard.py ← Streamlit UI
│ └── docs/ ← Logo/architecture assets
│
├── memv/ ← github.com/vstorm-co/memv
│ ├── README.md
│ ├── CHANGELOG.md
│ ├── pyproject.toml ← PyPI: memvee, v0.1.0
│ ├── docs/ ← docs site (MkDocs)
│ ├── src/
│ │ └── memv/ ← segmentation, extraction, validity, retrieval, storage
│ └── tests/
│
├── supermemory/ ← github.com/supermemoryai/supermemory (lean subset: schemas, SDK, MCP, arch docs)
│ ├── LICENSE
│ ├── README.md ← provenance + open-source vs hosted-backend split
│ ├── packages/
│ │ ├── validation/ ← Zod schemas (data model definitions)
│ │ │ ├── schemas.ts
│ │ │ └── api.ts
│ │ ├── lib/
│ │ │ ├── api.ts ← reveals backend dependency (api.supermemory.ai)
│ │ │ └── similarity.ts ← client-side cosine sim (visualization only)
│ │ └── tools/src/shared/
│ │ └── memory-client.ts ← SDK client (profile search, prompt formatting)
│ ├── apps/mcp/src/
│ │ └── server.ts ← MCP server (memory/recall/whoAmI tools)
│ └── skills/supermemory/references/
│ └── architecture.md ← claimed design (558 lines)
│
└── clawvault/ ← github.com/Versatly/clawvault
├── README.md
├── PLAN.md ← issue #4: ledger, reflect, replay, archive
├── CHANGELOG.md
├── SKILL.md
├── package.json ← npm: clawvault, v2.6.1
├── src/
│ ├── commands/ ← archive, context, inject, observe, reflect, replay, wake, sleep, task, project, ...
│ ├── observer/ ← compressor, reflector, router, session-watcher
│ ├── lib/ ← vault, memory-graph, ledger, observation-format, session-utils
│ └── cli/
├── bin/ ← CLI entry + command registration modules
├── hooks/ ← OpenClaw hook handler
├── dashboard/ ← web dashboard (vault parser, graph diff)
├── schemas/
├── scripts/
├── templates/
└── tests/
- Phased build order matters: Core memory first (write/read/decay), reliability second (dedup/maintenance/recovery), intelligence last (graphs/trust/cross-agent). Building out of order amplifies flaws.
- Tiered retrieval: Summary files first (fast, cheap), vector search fallback (thorough, expensive). Don't vector-search everything.
- Score decay:
final_score = relevance × exp(-λ × days)— recency-weighted relevance is universal across all architectures. - Feedback loops: Echo/fizzle (track which injected memories get used), behavior loops (extract corrections as lessons), learning loops (convert expensive LLM checks into cheap static rules).
- SQLite over hosted vector DBs: At current scales (1K-5K entries), SQLite + FTS5 + local embeddings outperforms hosted solutions on latency, cost, and operational simplicity.
- Multi-agent convergence: Shared memory creates homogenization pressure. Workspace isolation + file routing guards help but don't fully solve it.
- Vault index pattern: Single scannable manifest (one-line descriptions) → load individual entries on demand. One file read instead of N.
