Trace-learning

Type: types/tag-readme.md

This tag gathers external systems that learn from their own agent traces: memory extraction from sessions, session capture and transcript mining, skill libraries built from past runs, and self-evolving prompts or harnesses. The defining comparison is trace-learning techniques in related systems. The pattern is a two-stage loop: raw traces accumulate as episodes, logs, or transcripts, then a distillation step, automatic or manual, produces something that changes later behavior: a memory entry, a rule, a prompt, a route, or a fine-tune. The tag marks system reviews in agent-memory-systems whose analysis found that loop, read from code for most reviews and from documentation for the five lightweight ones; it is assigned by the review type's rule, not by a system's own claims, and only when the write side is fed by traces and a distillation mechanism exists. Nearby but different: agent-memory covers memory architecture whether or not traces feed it. About two thirds of the reviews carry the tag, so this head is selective; the scoped rg over kb/agent-memory-systems/ recovers the full set.

Start here

Representative systems

  • HyperAgents — benchmark feedback promotes executable changes to the harness itself
  • auto-harness — a benchmark-driven loop mines training traces and evolves the agent program
  • Dynamic Cheatsheet — test-time prompt memory curated by the model from its own solutions
  • REM — episodic memory service built from trace-learning episodes with vector and graph retrieval
  • Synapptic — mines transcripts into weighted user profiles
  • memwiki — agent-maintained trace notes in a hot-cache wiki
  • agent-memory — the memory architecture the distilled stage lands in
  • self-improving-systems — the theory of systems that change their own organization; trace-learning systems are its largest external casebook
  • deploy-time-learning — the post-release change these loops are built to absorb

Other tagged notes

  • ACE - ACE review: trace-learning playbook evolution, reflector-scored bullets, curator additions, optional deduplication, and coarse prompt read-back
  • AgeMem - Lightweight doc-grounded coverage of AgeMem — an RL-trained LTM/STM memory-management policy known from its paper, not from inspected code
  • Agent Skills for Context Engineering - Agent Skills for Context Engineering review: authored context-engineering skills plus a file-based researcher OS and trace-to-skill example tooling
  • Agent Workflow Memory - Agent Workflow Memory review: web-agent workflow files induced from successful traces and pushed into WebArena/Mind2Web prompts
  • Agent-R - Agent-R review: MCTS trace collection, revision-trajectory synthesis, checkpoint-level read-back, and no runtime retrieval store
  • Agent-S - Agent-S review: GUI agent framework with S1/S2 JSON experience memory, embedding retrieval, S3 reflection, and BBON trace evaluation
  • AgentFly - AgentFly/Memento review: planner-executor agent with JSONL case-bank memory, trace-judged case writes, and parametric or SimCSE case read-back
  • Agentic Harness Engineering - Agentic Harness Engineering review: trace-driven outer loop that distills coding-agent rollouts into debugger reports and durable harness edits
  • Agentic Local Brain - Agentic Local Brain review: local PKM capture into Markdown, SQLite, Chroma vectors, mining tables, RAG chat traces, and recommendation ranking
  • AI-Context-OS - AI-Context-OS review: filesystem-first Markdown memory with L0/L1/L2 context loading, generated adapters, MCP/chat read-back, and trace-learning optimization suggestions
  • ai-memex-cli - ai-memex-cli review: Git-backed Markdown vault, agent skill workflows, trace distillation, lint/watch loops, and context bootstrap
  • ai-modules - ai-modules review: deployable multi-vendor skill/plugin bundle with Markdown wiki, session wrapup, task backlog, and linted file memory
  • Amazon Science SAGE - Amazon SAGE review: AppWorld rollouts become reusable Python skills, retrieval state, SFT data, and GRPO reward signal
  • AriGraph - AriGraph review: in-run knowledge-graph world model with episodic observation memory, Contriever retrieval, LLM extraction/refinement, and prompt pushback
  • Ars Contexta - Ars Contexta review: Claude Code plugin deriving file-based agent knowledge systems with generated context, skills, hooks, trace mining, and coarse push read-back
  • Auto-claude-code-research-in-sleep - ARIS review: Markdown skill harness for autonomous research with project research-wiki memory, review traces, and gated trace-learning skill optimization
  • Autocontext - Autocontext review: iterative evaluation harness with trace-learning playbooks, hints, skills, tools, validators, runtime traces, and optional model distillation
  • Basic Memory - Basic Memory review: local-first Markdown knowledge graph with SQLite/Postgres indexes, MCP pull tools, semantic search, schemas, and Claude hook read-back
  • Beever Atlas - Beever Atlas review: chat-ingestion knowledge base with Weaviate facts, Neo4j graph memory, MongoDB wiki pages, MCP retrieval, and trace-learning wiki synthesis
  • browzy.ai - browzy.ai review: terminal personal KB with Markdown/wiki files, SQLite FTS, LLM compilation, query-time context assembly, and trace-learning digests
  • byterover-cli - byterover-cli review: local context-tree memory with HTML topic curation, BM25/runtime-signal retrieval, MCP hooks, review logs, dream pruning, and ByteRover cloud sync
  • cass_memory_system - cass-memory review: file-backed procedural memory for coding agents with cass session search, diary summaries, LLM reflection, scored playbook rules, MCP tools, and trauma guards
  • Claude Context Guard - Claude Context Guard review: Claude Code slash-command memory using project safeguard files, audits, pagination, hooks, and itemised code indexes
  • Claude Workstream Kit - Claude Workstream Kit review: repo-local active-work memory for Claude Code with ACTIVE.md resume, workstream closure, gates, and verifier agents
  • claude-obsidian - claude-obsidian review: Obsidian vault memory with agent skills, hot cache, wiki ingestion, hybrid retrieval, locking, hooks, and methodology modes
  • ClawVault - ClawVault review: deprecated markdown vault memory with graph/search context, OpenClaw prompt hooks, observer compression, facts, and maintenance workers
  • Clude - Clude review: cognitive memory SDK and MCP server with SQLite/Supabase stores, hybrid recall, dream-cycle synthesis, memory packs, and prompt-file push surfaces
  • Cognee - Cognee review: graph/vector agent memory control plane with session cache, recall routing, trace-learning improve loops, MCP tools, and decorator push
  • Compound Engineering Plugin - Compound Engineering review: repo-file workflow memory with generated strategy, brainstorm, plan, solution, pulse, session-history, and review artifacts
  • Continuity - Continuity review: local-first desktop AI workspace with shared SQLite memory, MCP tools, narrative synthesis, prompt push, and org sync
  • CORAL - CORAL review: filesystem multi-agent coding hub with shared notes, skills, attempts, roles, eval feedback, heartbeat prompts, and worktree isolation
  • Cortex - Cortex review: local RDF/SQLite cognitive knowledge service with ontology, hybrid retrieval, MCP tools, reasoning, and access-derived tier learning
  • cq - cq review: Mozilla AI plugin and MCP store for structured agent knowledge units, review-gated sharing, and agent-led reflection
  • CrewAI Memory - CrewAI Memory review: unified vector memory with LLM extraction, scoped recall, task/HITL learning, tools, and pre-task prompt injection
  • Decapod - Decapod review: Rust repo-native governance kernel with SQLite stores, context capsules, trace lessons, proof gates, and pull-first memory reads
  • deja-vu - deja-vu review: local lexical memory over the session stores of 22 coding-agent harnesses (JSONL and SQLite) with redacted file index, one MCP tool with modes, hooks at session start, on each prompt and around tool calls, a curated note layer, and sync/share
  • dense-mem - Dense-Mem review: self-hosted MCP memory server with Neo4j evidence, typed claims, verifier gates, fact promotion, and tiered recall
  • DocMason - DocMason review: repo-native private-document KB with provenance, governed ask, deterministic retrieval, and interaction-memory promotion
  • EchoesVault / echoes-vault-opencode - EchoesVault review: OpenCode plugin that bootstraps a Markdown/Obsidian vault, slash-command read-back, and agent-mediated trace capture
  • Eidetic - Eidetic review: Claude Code Markdown memory with hook-pushed context, FTS/vector recall, trace capture, compounding, drift penalties, and vault export
  • EQUIPA - EQUIPA review: SQLite-backed agent orchestrator with trace-learning lessons, episodes, prompt variants, and prompt-time read-back
  • ExpeL - ExpeL review: trace-learning benchmark agent that distills task trajectories into rules and retrieves prior trials as few-shots
  • G-Memory - G-Memory review: trace-learning multi-agent memory with Chroma task storage, NetworkX task graph, JSON insights, and orchestrator-pushed examples/rules
  • GBrain - GBrain review: Postgres/PGLite-backed agent brain with markdown write-through, hybrid retrieval, graph links, hot facts, skills, and dream-cycle maintenance
  • Graphiti - Graphiti review: temporal graph memory with episode provenance, LLM extraction, fact invalidation, hybrid retrieval, MCP tools, and pull-only activation
  • HALO - HALO review: trace-learning agent-harness optimizer with SQLite desktop trace store, JSONL trace indexes, recursive trace agents, and local analysis runs
  • Hermes Agent - Code-grounded review of Hermes Agent's bounded prompt memory, progressive skills, session recall, background trace learning, and skill curation
  • Hindsight - Hindsight review: service-backed agent memory with LLM fact extraction, observations, hybrid recall, integrations, hooks, transfer, and trace-learning
  • Incremental Self-Improvement - Lightweight doc-grounded coverage of Schmidhuber's incremental self-improvement paradigm, a reward-gated policy self-modification system
  • Kompl - Kompl review: SQLite-backed knowledge compiler that ingests sources into a generated wiki with provenance, FTS/vector retrieval, MCP tools, and chat-derived drafts
  • LACP - LACP review: local control-plane agent harness with trace-learning Obsidian/SMS memory, hook-time context injection, RAG pull, and policy gates
  • Letta - Letta review: stateful agent server with core memory blocks, archival and recall tools, compaction, sleeptime memory agents, and optional git-backed memory
  • Link - Link review: local Markdown wiki memory with raw captures, reviewed memory pages, bounded query packets, MCP/CLI skills, validation, and local viewer
  • LLM Wiki (kenhuangus) - LLM Wiki review: local-first Obsidian/wiki compiler with source monitors, LLM extraction and integration, BM25 search, autonomous maintenance loops, and weak trace-learning prompt-optimization scaffolding
  • LLM Wiki (nvk) - LLM Wiki review: portable agent plugin that compiles source files into topic wikis, queryable through index-guided reads, audits, linting, and session lessons
  • llm-project-wiki - llm-project-wiki review: prompt-only Claude Code workflow that bootstraps an Obsidian project wiki, wiki-first rules, diff ingest, and gap audits
  • llm-wiki (Pratiyush) - Pratiyush llm-wiki review: local file-based transcript-to-wiki compiler with raw session capture, agent-authored wiki pages, static AI exports, MCP tools, and optional synthesis
  • LLM-WIKI-MCP - LLM-WIKI-MCP review: local Markdown wiki with SQLite FTS, MCP/CLI retrieval, Ollama ask, provenance ingest, sidecar notes, and ask-history memory
  • Mem0 - Mem0 review: memory SDK/server/platform with additive trace extraction, hybrid retrieval, agent plugins, hooks, and pushed context injection
  • Memori - Memori review: SDK and agent integrations with trace-learning augmentation, SQL/Rust storage, hybrid recall, and pre-call memory injection
  • MemoryOS - MemoryOS review: hierarchical conversational memory with trace-learning summaries, profiles, knowledge extraction, vector retrieval, and pre-call prompt assembly
  • MemPalace - MemPalace review: local-first ChromaDB/SQLite memory palace with transcript mining, MCP tools, hooks, and explicit wake-up/search read-back
  • MentisDB - MentisDB review: append-only hash-chained agent memory with MCP/REST tools, skill registry, LLM extraction, and LangChain memory
  • Meta-Harness - Meta-Harness review: trace-learning harness search that uses logs, evaluations, proposer skills, and generated code to evolve memory and agent scaffolds
  • MiroShark - MiroShark review: simulation knowledge graph with Neo4j graph memory, trace-learning agent activity edges, report-agent reasoning traces, and many public export surfaces
  • Mnemosyne / IsaacCLupus mnemosyn spec - Lightweight doc-grounded coverage of Mnemosyne, a spec-first local semantic memory OS with unified SQLite, vault, MCP, and agent-pack design
  • nao - nao review: analytics-agent context builder with file-backed project context, SQL guardrails, stories, and pushed trace-learning user memory
  • Neo4j Agent Memory Service (NAMS) - Doc-grounded review of NAMS skill distillation and governance: scoped context-graph memory becomes provenance-linked, review-gated procedure artifacts with drift repair
  • Nuggets - Nuggets review: TypeScript HRR fact memory, Pi prompt injection, trace capture, hit-count promotion to MEMORY.md, and Telegram/WhatsApp gateway
  • OpenSage - OpenSage review: ADK agent framework with dynamic subagents, Skills, sandbox memory, Neo4j history/memory, plugins, and RL adapters
  • OpenViking - OpenViking review: context database with viking:// files, session-derived memory, hierarchical retrieval, hooks, MCP, and LangGraph injection
  • Origin - Origin review: local AI-work memory daemon with sourced pages, hybrid retrieval, review gates, git-backed Markdown, and MCP/Claude Code read-back
  • OS-Copilot - OS-Copilot review: FRIDAY promotes judged Python execution traces into Chroma-retrieved reusable tools for later planning and codegen
  • Phantom - Phantom review: VM co-worker with Qdrant memory, heuristic session extraction, and queued self-evolution over config files
  • pi-self-learning - pi-self-learning review: pi extension that reflects completed agent sessions into git-backed daily, core, and long-term memory files
  • Pond - Pond review: Lance-backed cross-client session archive with canonical codecs, scheduled trace acquisition, pull-only recall, read-only analytics, and restore
  • ReasoningBank - ReasoningBank review: trace-learning benchmark memories selected by embeddings and injected into WebArena and mini-SWE-agent prompts
  • Reflexion - Reflexion review: benchmark agents turn failed trajectories and test feedback into task-local verbal lessons for later attempts
  • SAGE - SAGE review: consensus-governed local agent memory with MCP turn capture, hooks, hybrid recall, decay, and corroboration
  • sage-wiki - sage-wiki review: LLM-compiled wiki memory with SQLite search/vector/ontology state, MCP pull tools, session capture, and trust gates
  • Scroll - Scroll review: SQLite-backed executable context manager with write-through interaction traces, tiered eviction maps, validated summaries, and programmable recall
  • Self-Training-LLM - Self-Training-LLM review: offline synthetic Wikipedia QA generation, uncertainty-filtered SFT/DPO datasets, and model-weight learning rather than contextual memory
  • Signet AI - Signet AI review: local-first daemon memory with SQLite, FTS/vector/graph recall, hook injection, transcript lineage, and guarded repair paths
  • SkillNote - SkillNote review: self-hosted SKILL.md registry with collections, imports, sync adapters, usage/rating feedback, and prompt-derived draft candidates
  • SkillRL - SkillRL review: trajectory-derived SkillBank JSON, prompt-time skill push, dynamic failed-trajectory updates, and SFT/RL policy learning
  • SkillWeaver - SkillWeaver review: web-agent trajectories distilled into Playwright API skills with LLM relevance push and verification metadata
  • SkillX - SkillX review: trajectory-derived planning, functional, and atomic skill libraries with filtering, merging, and prompt-time retrieval
  • Smriti-MCP - Smriti-MCP review: MCP markdown memory server with file-backed notes, lexical recall, traces, salience, wikilinks, and agent-mediated consolidation
  • Spacebot - Spacebot review: Rust team-agent harness with SQLite graph memory, LanceDB hybrid recall, pushed working context, trace-learning persistence, and skill injection
  • supermemory - Supermemory review: hosted memory API with generated SDK contracts, profile/search injection middleware, MCP tools, browser capture, graph UI, and trace-learning memory
  • Synto - Synto review: local LLM vault compiler that turns raw notes into reviewed Markdown wiki articles, SQLite identity state, agent packs, and MCP read tools
  • TheKnowledge - TheKnowledge review: file-first LLM wiki gateway with citation-grounded Markdown, NotebookLM synthesis, MCP tools, and policy distillation
  • Trajectory-Informed Memory Generation - Lightweight doc-grounded coverage of Trajectory-Informed Memory Generation — an IBM trajectory-to-tip pipeline known from its paper, not inspected code
  • Virtual Context - Virtual Context review: proxy-owned context virtualization with trace-learning compaction, facts, paging tools, and prompt-time memory injection
  • Voyager - Voyager review: Minecraft lifelong-learning agent with trace-learning executable skill libraries, Chroma retrieval, curriculum QA cache, and prompt pushback
  • WeKnora - WeKnora review: enterprise RAG and agent platform with document chunks, wiki pages, graph memory, ReAct tools, and push plus pull read-back
  • WUPHF - WUPHF review: local multi-agent office with git-backed markdown wiki, per-agent notebooks, fact extraction, learning logs, lint, and cited lookup
  • xMemory - xMemory review: trace-learning hierarchical agent memory with JSONL stores, Chroma/BM25 search, semantic themes, graph files, and pull-only retrieval
  • Zikkaron - Zikkaron review: Claude Code MCP memory with SQLite/FTS/vector storage, predictive write gating, trace-learning consolidation, hooks, and push/pull recall