NEW RUNTIME

Topic hub

Agent memory

Evidence on agent memory as a write loop: what should persist, who can correct it, and how it changes future work.

Short answer

  • Agent memory is not only retrieval; it is a governed write loop that decides what survives a task.
  • Useful memory needs correction-time updates, provenance, decay, and behavioral evaluation.
  • The failure mode is treating memory as a passive store while the actual system behavior drifts.

Pattern memory

Current hypotheses

1 patterns
  1. high

    Company memory needs write and correction loops

    Company memory becomes useful when real work continuously updates a governed knowledge layer, not when an agent merely searches a larger document pile.

Field notes

Longer analysis

1 notes

  1. Cognee Packages Agent Memory as a Self-Hosted Knowledge Graph

    Cognee matters less as another memory SDK and more as an attempt to package persistent agent memory through ingestion, graph/vector search, ontology, and self-hosting.

Raw signals

Recent observations

9 signals

  1. Agent memory is being evaluated as system behavior

    MemoryData evaluates what an agent stores, retrieves, updates, and forgets across a sequence instead of grading one final answer.

  2. Codebase Memory Combines Search with a Repository Map

    A code retrieval tool adds structural repository context to text search so an agent can locate relevant code without loading the whole project into context.

  3. Agent Memory Needs Its Own Review Loop

    Persistent memory can amplify mistakes across future tasks, so edits, provenance, and correction paths need explicit operator control.

  4. Deep Agents combines plans, files, memory, and subagents

    The open-source Deep Agents stack packages common long-horizon primitives on top of LangGraph and LangChain.

  5. DoorDash makes memory a versioned personalization layer

    DoorDash builds structured memory blocks from behavior and stores model, schema, timestamp, prompt, and response lineage for each component.

  6. OpenHuman Builds a Personal Harness Around Local Memory

    OpenHuman combines personal data, memory, tools, and an operator layer into a durable system instead of another isolated assistant chat.

  7. AgentMemory Shares State Across Codex, Claude Code, and Cursor

    A common memory layer reduces the cost of switching coding clients and moves context ownership outside any one model interface.

  8. OpenAI Shows Memory Compaction for Long-Running Agents

    The Agents SDK pattern treats compaction and state transfer as explicit reliability mechanisms for work that outlives one context window.

  9. CocoIndex Keeps Agent Context Fresh Without Full Reindexing

    CocoIndex turns source changes into incremental dataflow updates for live agent context and code-aware retrieval.