Memory Should Be Auditable
Most memory stacks behave like mutable caches: new facts overwrite old ones and the system keeps only a single view of reality. That design is convenient, but it destroys history - exactly the thing long-running agents need to answer questions like "When did this change?" or "Why was this decision made?"
At Crosmos we treat memory as an audit log. Current state is derived from history, never written over it. That's the foundation of our Monotonic Temporal Knowledge Graph (MTKG).
We never rewrite history
Every interaction is converted into structured entities and relationships, then appended to the graph with timestamps and provenance. If a user switches jobs, the previous role remains. If a team reverts an architecture decision, both decisions coexist with their temporal context.
Because the records are append-only, we always know:
- what changed
- when it changed
- why it changed (the evidence sits in the originating memory)
State is resolved at retrieval
Preserving history introduces a harder problem: deciding which fact is true right now. Agents often need to pick between multiple valid statements like:
- User worked at OpenAI
- User joined Anthropic
- User advises a startup
Instead of mutating data at ingest, we resolve state at query time. When a request arrives, Crosmos fans out across four parallel signals:
- semantic similarity
- keyword search
- graph traversal
- temporal reasoning
Candidates from each signal are fused (Reciprocal Rank Fusion), rescored for recency and temporal proximity, and finally re-ranked with a cross-encoder. The answer is computed from history every time.
Graph is projection, not truth
The graph is a powerful projection for traversal, but the source of truth is the underlying memory record - the episodic data, extracted facts, viewpoints, timestamps, and metadata captured exactly as they were observed. Every node/edge carries provenance back to those records, and each record links back to its original source (Slack thread, email, doc, call notes, etc.).
Source → Memory → Graph
That chain gives agents explainability: follow any retrieved item back to the raw evidence.
Relationships evolve
Relationships are temporal objects. Each edge stores:
- provenance to the originating memory
- extraction confidence
- temporal metadata describing when it became valid
Scoring during traversal favors higher-confidence, more recent edges without deleting older context. Historical queries can still surface prior states, while present-tense queries remain accurate.
Memory as infrastructure
Enterprise agents need more than convenience - they need auditable context. Crosmos delivers that by combining:
- episodic memory (append-only log)
- MTKG for structured reasoning
- provenance tracking for every hop
- state-aware retrieval with temporal weighting
When an agent surfaces context, you can trace it end-to-end:
Source → Memory → Graph → Retrieval
That makes memory trustworthy. Instead of hand-waving around stale caches, you get a system that records how knowledge evolves, maintains temporal awareness, and keeps every decision explainable.
Ready to build on top of MTKG? Grab an API key at crosmos.dev and join the conversation on Discord.