Back to Gap Map Product / UX / Agent memory & context stores

agentmemory

Rohit Ghumare
source available / Overall score: 3.2

agentmemory records what coding agents do through hooks and plugins for tools such as Claude Code, Codex and Cursor, condenses those observations into session summaries, facts and lessons, and serves them back to later sessions through keyword, vector and graph search. It runs as a local server with MCP and REST interfaces and a live viewer, built on the iii engine. Rohit Ghumare maintains it.

Openness

2 medium confidence
2.0
license
Apache-2.0 (agentmemory's server, CLI, MCP shim and plugins)+Elastic-License-2.0 (the iii engine, the runtime the agentmemory CLI downloads and starts by default)
source
public (github.com/rohitg00/agentmemory, with the iii engine source at github.com/iii-hq/iii)
core features withheld
no — the 1,169-path tree at 5c9e2e6 carries the root Apache-2.0 LICENSE and identical LICENSEs in packages/mcp and integrations/pi, with no ee/, enterprise/, commercial/ or proprietary/ path; no paid tier is offered

agentmemory's own code is Apache-2.0 licensed and published in full, with nothing held back for a paid tier. It runs on the iii engine, which its command-line tool downloads and starts automatically, and that engine is under the Elastic License 2.0, which forbids offering it to others as a hosted service.

Adoption

2 medium confidence
2.0

Adoption is read from monthly npm downloads of @agentmemory/agentmemory, the package the setup command runs. The separate @agentmemory/mcp shim is left out because every install of it also pulls the main package, so counting both would count one install twice.

Capability

5 medium confidence
5.0

agentmemory captures coding-agent activity itself, condenses it with a language model into summaries, facts and procedures, supersedes outdated versions and lets unused memories decay and expire. Its extraction and consolidation run only once a model provider is configured; without one it stores observations and recalls them by keyword.

  • https://raw.githubusercontent.com/rohitg00/agentmemory/main/README.md recorded 2026-11-20

    The README's memory pipeline runs session summaries, knowledge-graph extraction and slot reflection at session end; its consolidation tiers are working, episodic, semantic and procedural, and "Memories decay over time (Ebbinghaus curve)... Stale memories auto-evict. Contradictions are detected and resolved."

  • https://www.agent-memory.dev/docs recorded 2026-11-20

    The documentation home calls AgentMemory "a local memory server for coding agents" that "records what your agent does, keeps the decisions and lessons that matter, and gives them back in the next session."

Verified 2026-11-20, except some scores, still awaiting re-confirmation