Agents lose everything between sessions — decisions, context, learned preferences. Mnemo gives them structured, persistent memory they can query, update, and share.
LoCoMo is the standard evaluation for long-conversation memory systems. Mnemo ranks second overall and first in multi-hop reasoning — the hardest category, requiring synthesis across multiple stored memories.
Mnemo decomposes conversations into typed memory atoms — episodic, semantic, procedural — each with Bayesian confidence that evolves as evidence accumulates.
Store structured knowledge with automatic decomposition. Bayesian confidence scoring means memories strengthen or weaken with repeated evidence — not just stack up.
Semantic search ranked by composite similarity and confidence. Temporal anchoring, domain tags, and tunable verbosity give agents precise control over what they retrieve.
Agent-to-agent memory sharing with explicit capability controls. The owner decides what gets shared and with whom. No blanket access, no data leaks.
A decision made in a chat session is available in a coding session. Preferences set in one conversation carry forward to the next. One agent identity, one memory, regardless of interface.
One team member's agent shares relevant context with another's — across operators, across time. Agents coordinate without real-time communication channels.
Direct it explicitly: "remember this decision." Let it ask: "should I store this?" Or let it learn on its own. You control how much autonomy the agent has over its own memory.
Agents that remember get better over time. Learned patterns, past mistakes, accumulated expertise — memory turns a stateless tool into a colleague.
Mnemo is source-available and in active development. We're working with a small group of design partners building production agent systems.