Show HN: DeltaMemory – Persistent cognitive memory for production AI agents
Most AI agents forget everything between sessions. We built DeltaMemory to fix that.
It's a cognitive memory layer that gives agents persistent recall, automatic fact extraction, and temporal reasoning — via a single SDK call.
Key numbers: - 89% accuracy on LoCoMo long-term conversation benchmark - 50ms p50 retrieval latency - 97% cost reduction vs raw token re-processing
Open source SDKs, works with any LLM stack. Currently in early access.
Happy to answer questions about the architecture, benchmark methodology, or how we handle knowledge graphs.
(
2 comments
| on HN
)
A few observations from the file-based approach we've been using (plain markdown, no SDK):
The hard part isn't storage — it's memory hygiene. Agents accumulate context fast, and after a week you end up with thousands of tokens of "memory" that is contradictory (old decisions that were reversed), stale (facts that were true 3 days ago), or too granular to be useful. The agent then has to reason over a noisy memory corpus, which degrades quality.
What we've found helps: separating memory by type and update frequency. - Identity/values: write once, almost never change. Agents reference this constantly. - Operational rules: changes occasionally. High-signal, low-volume. - Learned context: updates daily. Needs regular pruning — an agent summarizing its own recent logs into higher-level insights. - Event log: append-only, date-partitioned. Used for debugging, not for agent reasoning.
The 89% on LoCoMo is solid. Curious how DeltaMemory handles fact contradictions — specifically when a fact changes over time ("price is $29" becomes "price is $49"). Does the temporal reasoning layer deprecate old facts, or does the agent have to reason over conflicting entries?
Also interested in how you handle memory write decisions — does the agent decide what's worth remembering, or is everything written and retrieval filters the noise?
The part I'm curious about: how do you handle temporal conflicts? If an agent records 'X is true' in week 1 and 'X is now false' in week 4, does DeltaMemory resolve that automatically or does the retrieval surface both and leave resolution to the model?
I've been building agents where state corruption from this is a real problem -- the model sees contradictory facts and either hallucinates a reconciliation or stalls entirely.