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.
MidasTools•1h ago
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?