AMA skill
A memory skill that stores turns, facts and episodes, verifies recall and refreshes stale entries, with an OpenClaw plugin.
- 115 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- User turns are stored in a local database and sent to external model and embedding endpoints, and memory may hold personal data.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/amaDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- User turns are stored in a local database and sent to external model and embedding endpoints, and memory may hold personal data.
This is third-party code. Review the repository files before installing.
What it does
AMA is a research memory system from an ACL 2026 paper and a ready skill for agents. Four roles run in a loop: a constructor stores raw turns, facts and episodes, a retriever rewrites the query and picks the right granularity, a judge checks relevance and requests another bounded retrieval round when evidence is thin, and a refresher removes or updates entries that are stale or conflict with newer ones. `SKILL.md` gives the agent policy: when to recall prior context, what to capture, how to synthesize episodes and how to honor a request to forget. The repository includes a reference adapter for OpenClaw with automatic recall and capture, CLI commands and a managed SQLite and FAISS sidecar, plus the original Python code for reproducing LoCoMo and LongMemEval experiments.
Who it is for. Builders of agents with long-term memory who want recall that is verified and kept current rather than just stored.
Good fit when
- The agent must remember that the user changed job or address and stop quoting the old one.
- An answer about old sessions needs an exact quote and date.
- A researcher reproduces results on LoCoMo and LongMemEval.
- The user asks to forget a specific fact and the entry must be deleted.
- A memory plugin is added to OpenClaw with automatic recall.
Not a fit when
- Simple one-file memory without Python or servers is enough.
- No API access to a chat model and embeddings is available.
- A ready MCP server is expected, while this offers an OpenClaw plugin and agent policy.
Example request
Remember that we moved to Postgres 17, and in the next conversation answer the database version question from stored memory.Limitations
It needs Python 3.10, Node.js 22.22.3 or newer (24.15 is also supported), pnpm 11 and two OpenAI-compatible endpoints: a chat model and embeddings. The API key is set through environment variables. A ready adapter exists only for OpenClaw; for other harnesses the repository offers `SKILL.md` as a portable policy and you write the integration yourself. Benchmark datasets are not redistributed. The authors admit temporal reasoning remains a weak spot. The last commit is from September 2026.
How to disable. Disable the `openclaw-ama` plugin in the OpenClaw plugin manager. Delete the SQLite database and FAISS index holding stored turns separately if memory is no longer needed.
FAQ
Which agents are supported?
The ready plugin targets OpenClaw. For other harnesses the agent follows `SKILL.md` and you must wire the storage layer yourself.
Which memory forms are stored?
Three: raw text with exact wording, atomic facts and episodes synthesized from several turns. The retriever chooses the level that fits the question.
How do I verify the installation?
Run `openclaw ama doctor`, the diagnostic named in the README. A direct check is `curl http://127.0.0.1:8321/health`.
Where does data go?
Turns and queries are sent to the chat-model and embedding endpoints you configure, while stored memory lives locally in SQLite and FAISS.
What license does it use?
Apache-2.0. The research code is in `python/AdaptiveMemory` and the OpenClaw adapter is at the repository root.
Discussion
No questions or reviews yet.
Related
SkillsHook and skill for OpenClaw: reorganize the memory and knowledge folders into L0/L1/L2 layers with a relation graph
A long-term memory plugin for the OpenClaw agent on LanceDB: captures preferences and decisions itself, hybrid search, noise fading
Long-term memory for AI agents: a skill and the mindmemos command store facts from conversations and recall them in later sessions
A research project that improves and deduplicates an agent's skill library from its sessions through a local proxy and an optional server
Self-Improving Agent
self-improving-agent
Skill for OpenClaw: the agent logs errors, corrections and missing features to .learnings files and promotes verified lessons into workspace memory
A Datalog engine in Rust as an agent's working memory: facts with provenance, inference rules and an MCP server for Claude Code
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| AMA (this card) | no | 115 | |
| Mem0 | no | 68,887 | a general memory layer for agents; AMA adds a verifying judge and refresh of outdated facts |
| Claude-Mem | no | 116,457 | session memory for Claude Code; AMA targets OpenClaw and open research |
README in short
In the paper AMA scores 0.774 on LoCoMo (GPT-4o-mini) against 0.717 for full context while using fewer input tokens. The method splits memory into three levels and has a judge verify recall.