Most software has the memory of a goldfish. This one doesn’t.
Your agents are brilliant and amnesiac. Every session starts from zero. memkeeper gives them a persistent memory to write to and recall from: hybrid semantic + keyword retrieval, reranked for relevance, and local-first.
“Remind me again, what did we decide last week?”
Context windows fill and flush. Useful facts get re-explained, re-pasted, re-paid for. The agent is sharp but starts every conversation a stranger.
“You decided to roll back past 2% errors. Want me to set the alert?”
Write a fact once; recall it for years. The right memory surfaces on its own. No re-prompting, no copy-paste, no forgetting.
See it recall.
Pick a question · watch hybrid rankingDense vectors catch meaning, BM25 catches exact terms, and a cross-encoder reranks, so the best memory lands on top, not just the closest keyword match.
Every memory, kept and findable.
Durable facts, decisions, and preferences, not transcripts. memkeeper indexes each memory three ways so the answer you need is one query away, years later.
Write once
Durable facts, decisions, and preferences, not transcripts. Concise by design.
Find by meaning
Semantic search that understands intent, with a keyword safety net for exact matches.
Surface the right one
A cross-encoder reranker reorders candidates so the best memory lands on top.
Measured, not asserted.
LoCoMo · reproducibleFull LoCoMo (1,982 evidence questions), out-of-the-box semantic + cross-encoder rerank. Warm serve daemon, ~32× faster than cold per-call. An optional ColBERT late-interaction pass lifts hit@20 to 0.894. Reproduce it yourself: scripts/memkeeper_locomo_benchmark.py · docs/benchmarks.md.
Stop starting from zero.
Get early access to memkeeper and give your agents a memory worth keeping.