Long-term memory for AI agents
Your users shouldn't have to repeat themselves
Mnemo remembers the conversations, decisions, and details your users already shared. When your agent needs them, it gets the right context with the source attached.
Mnemo is long-term memory infrastructure for AI agents: hybrid retrieval across seven signals (vector, BM25, atomic fact, temporal, entity graph, concept boost, semantic bridge) fused with intent-aware Reciprocal Rank Fusion, atomic fact extraction, and prompt-ready profiles. Model-agnostic and US-hosted. Free tier: 1K writes and 10K searches each month, no credit card.
Measured, not asserted
- ~150ms — warm search p50 on the production API.
- 99.2% — answer-session found in the top 5 on LongMemEval-S (496 of 500, Recall@5, top-50 candidate chunks, no LLM reranker).
- 85.2% — end-to-end answers judged correct on LongMemEval-S under the official LongMemEval judge.
Store it once. Find it when it matters.
Add a detail when it happens. Search for it later in the language your user actually uses.
Find the right memory. Not just a plausible one.
98.0% recall with only five chunks. Increasing the candidate pool tenfold moves recall by 1.2 points — the result saturates early instead of depending on an oversized search pool.
A larger context window is still not a memory.
Your agent does not need every old message. It needs the few past details that change what it should say or do next.
Chat history
Useful at first. Then the prompt grows, old decisions get compressed, and every request carries messages the agent does not need.
Document RAG
Great for manuals and knowledge bases. Less useful when a user changes their mind or a fact only makes sense in time.
Mnemo
Keeps the details worth remembering, retrieves only what matters now, and returns the source that supports each result.
From a conversation to context your agent can use
I. Send the history
Send a conversation, document, or transcript through the SDK or API.
II. Keep the details
Mnemo pulls out the facts, dates, people, and decisions that may matter later.
III. Ask naturally
Search the way a user speaks. Mnemo looks across meaning, wording, entities, and time.
IV. Check the source
Each result includes confidence and a reference to the conversation or document behind it.
Your users' memory, already filled in
Memory shouldn't start empty. Connect the tools your users already live in and Mnemo keeps their context continuously in sync — 20 sources, backfilled, updated, and deletion-aware, landing in the same search as everything else.
Built for the moments users notice
Good memory is quiet. The conversation simply continues from the right place.
- Customer support — pick up with the customer's setup, the unresolved issue, and the next step your team already promised.
- Personal assistants — remember preferences, plans, and changing circumstances without asking the same questions every session.
- Workflow agents — carry decisions and outcomes from one run into the next, with enough history to explain what changed.
The unglamorous details that make memory useful
- Find the detail, not just a similar paragraph — Mnemo searches meaning, wording, facts, entities, and time together, then combines the evidence into one ranked result.
- Keep small facts from disappearing — long conversations become short, reusable facts without losing the people, dates, or source they came from.
- Try again when the first search is weak — low-confidence searches can be rewritten and run again instead of quietly returning a poor match.
- Keep dates and changes straight — ask what happened last week, before a migration, or after a user changed their preference.
- Keep the model and framework you already use — Mnemo handles memory beside your existing agent.
- Know whose memory you are reading — workspace isolation, scoped API keys, audit history, and US data residency are part of the service, not application TODOs.
Use the tools already in your stack.
Seven packages published across npm and PyPI: getmnemo, getmnemo-cli, getmnemo-mcp, getmnemo-anthropic, getmnemo-vercel-ai, getmnemo-mastra, and the Python getmnemo on PyPI. Install the SDK with `npm install getmnemo`, the CLI with `npm install -g getmnemo-cli`, or run the MCP server with `npx -y getmnemo-mcp`.
Start the next conversation with context
Your users have already told you what matters. Give your agent a way to use it. Free tier: 1K writes and 10K searches each month, no card.