Separate memory by the job it does
Working context holds what a task needs now. A checkpoint records enough state to resume work. Episodic history records what happened. Semantic memory holds claims worth reusing. Procedural memory stores how to do something. These can share infrastructure while retaining different lifecycles and access rules.
For a research assistant, an unfinished search query belongs in working state. A source-backed naming convention may become semantic memory. A reviewed procedure for comparing announcements belongs in a skill. Saving every transcript into one searchable bucket blurs these distinctions and makes correction difficult.
Store why you believe a claim
A useful record has a stable identifier, claim, evidence reference, owner, observed time and review condition. Add a privacy class and a retention rule. If a policy changes, keep the relationship between the old and new records so a later reader can reconstruct why the answer changed.
The SIS work-graph architecture separates event history from promoted memory. Its memory-provider contract distinguishes a shared or remote provider from a heavyweight process created for every agent. Read these as concrete design artifacts. Do not infer that a documented provider is configured in your own environment.
Test recall, contradiction and forgetting
Ask whether retrieval returns the right evidence for a question, not merely a semantically similar paragraph. Plant two dated contradictory facts in a synthetic collection. Does the system identify their time scope, or blend them into one confident answer?
Then test deletion and expiry. A removed record may survive in caches or derived indexes unless the system propagates the change. Document what is canonical, what can be rebuilt and how forgetting is verified. A memory system becomes trustworthy through those observable behaviors, not through how many notes it holds.
Inspect a small knowledge graph
Follow the claim back to evidence.
Claim · release-date
The draft states a release date. Its identifier stays stable so evidence and revisions can refer to the same claim.
- Supported by
- source:announcement-a
- Included in
- artifact:digest-v1
- Confidence
- Requires source inspection
The arrows are a simplified reading path. The records expose the actual relationship names.
A worked example · Illustrative
A durable memory record
{
"id": "research.source-rule.v1",
"claim": "Material claims need a dated primary source",
"evidence": "public-method.md#source-policy",
"observedAt": "2026-09-15",
"owner": "practice-team",
"visibility": "public-example",
"reviewWhen": "source policy changes",
"supersedes": null
}Put the idea to work
Your practice
- Sort five example records into working, checkpoint, episodic, semantic or procedural memory.
- Write evidence, ownership and retention fields for one durable claim.
- Introduce a newer contradictory record and define how both remain interpretable.
- Describe a deletion test covering the canonical store and derived index.