Agentic Memory

New threads lose working agreements, style rules, and decisions unless explicitly rehydrated.

Each harness stores memory differently (or not at all). Cross-tool recall is manual copy-paste.

Naively dumping history into long windows degrades attention (lost-in-the-middle / RULER-class effects).

Vendor "memory" that cannot export, audit, or delete under user control is convenience without ownership.

Current turn state, scratchpads, tool results — ephemeral, high churn, aggressively compressed.

What happened: runs, decisions, handoffs, receipts with timestamps and actors.

Stable facts, preferences, entities, brand voice, architecture decisions — pocket-sized truths.

How to do work: skills, playbooks, hooks, SOPs that load on demand.

Policy for store/retrieve/update/summarize/discard — rules, heuristics, or RL-trained ops (AgeMem-class research).

Mem0-class systems: extract memories, embed, retrieve. Fast adoption, strong ecosystem, weaker pure temporal multi-hop unless graph options enabled.

Graphiti/Zep-class: facts as time-stamped relationships. Stronger on "what was true when" and multi-hop entity questions.

Letta/MemGPT lineage: treat context as RAM and external store as disk; agent edits memory blocks deliberately.

Starlight Memory doctrine: filesystem-native atoms with hybrid recall as authority; Mem0/Hindsight/etc. as scored adapters behind a provider contract.

Universal memory layer; largest community signal; Apache 2.0 core + managed cloud; strong personalization path; optional graph modes in later stacks.

Temporal context graphs; Graphiti OSS engine; production Zep platform for governance/retrieval; strong temporal reasoning reports on LongMemEval-class benches.

Stateful agents with self-editing memory; OS metaphor; strong for long-horizon agents that manage their own memory blocks.

LangGraph-native memory SDK for teams already on LangChain state machines.

Provider contract + local_core authority + MCP tools (recall/search/remember) for multi-harness coding fleets; evaluation-driven defaults.

Cognee (graph-first), Supermemory (MCP-first), Honcho/Hindsight-class systems, pure vector DBs (Pinecone/Weaviate/Qdrant) as storage not full memory OS.

Standard agent loop: retrieve relevant memory, act, then write durable deltas with provenance.

Prefer durable rules and atomic facts over raw chat dumps. Every persistent token taxes every future turn.

Keep procedural knowledge in skills loaded on demand; do not paste entire skill libraries into every prompt.

Track who said/did what across agents to prevent credit and contradiction collapse in swarms.

Anthropic-class techniques: compact history, keep structured notes outside the polluted transcript.

Local vault is canonical; cloud mirrors are derived. Deletion and retention must be enforceable.

Long-horizon chat memory abilities: extraction, multi-session reasoning, temporal updates, abstention.

Multi-session dialogue recall and multi-hop reasoning across long conversations.

Did the agent retrieve the right memory before acting? Did it write the decision after? Score intermediate steps.

Recall quality, contradiction rate, latency, cost, privacy, exportability — Starlight Memory observatory pattern.

You must be able to export, delete, and re-host memory atoms without vendor permission theater.

Route by classification: public, internal, private-life, regulated. Block cloud writes for forbidden classes.

Contractual and technical controls against using private life memory as training fuel.

Dozens of coding agents per machine: heavy memory providers must be shared daemons/APIs, not one runtime per terminal.

Skills and hooks load procedural memory; session start restores working agreements; stop hooks write durable lessons.

Domain packages (business, creator, health, family, investor) share SIP/SIS memory nodes without cross-contaminating private data.

starlight-evals-class harnesses test retrieval + agent trajectories, not only chat recall quizzes.

Handover protocols and ledgers make multi-session, multi-human, multi-agent continuity operational.