Memory Types

contextdb models cognitive memory types with different decay rates. This is primarily useful in agent_memory namespaces where different kinds of knowledge should fade at different speeds.

The four types

Type	Constant	Alpha	Half-life	What it models
Working	core.MemoryWorking	999.0	~seconds	Current task context, scratch state
Episodic	core.MemoryEpisodic	0.08	~8.7 hours	Specific events and interactions
Semantic	core.MemorySemantic	0.02	~34.7 hours	Generalized facts and knowledge
Procedural	core.MemoryProcedural	0.001	~29 days	Learned skills and workflows
General	core.MemoryGeneral	0.05	~13.9 hours	Untyped (default)

Using memory types

Set MemType on writes to control how fast a memory decays:

// This interaction happened just now: will fade in hours
ns.Write(ctx, client.WriteRequest{
    Content: "User asked about deployment",
    SourceID: "agent:self",
    Vector:   embedding,
    MemType:  core.MemoryEpisodic,
})

// This is a learned fact: will persist for days
ns.Write(ctx, client.WriteRequest{
    Content: "The production database is on us-east-1",
    SourceID: "agent:self",
    Vector:   embedding,
    MemType:  core.MemorySemantic,
})

// This is a learned procedure: will persist for weeks
ns.Write(ctx, client.WriteRequest{
    Content: "To deploy: build, push image, helm upgrade",
    SourceID: "agent:self",
    Vector:   embedding,
    MemType:  core.MemoryProcedural,
})

Decay mechanics

The decay function is:

recency_score = exp(-alpha * age_in_hours)

After one half-life, the recency score drops to 0.5. After two half-lives, it's 0.25. The memory doesn't disappear. It just ranks lower in retrieval results.

Hours elapsed	Working	Episodic	Semantic	Procedural
1	~0.00	0.92	0.98	1.00
8	~0.00	0.53	0.85	0.99
24	~0.00	0.15	0.62	0.98
168 (1 week)	~0.00	~0.00	0.03	0.85
720 (1 month)	~0.00	~0.00	~0.00	0.49

When to use which type

Working memory: For ephemeral context that's only relevant during the current task. Think "the user just said X" or "I'm currently working on Y". Drops off almost immediately.

Episodic memory: For specific interactions and events. "The deploy failed at 3pm" or "User asked to increase the timeout". Important for a few hours, then fades.

Semantic memory: For generalised facts extracted from episodes. "The API uses JWT auth" or "Deploys take about 5 minutes". Persists across sessions.

Procedural memory: For learned skills and workflows. "To fix a broken deploy: check logs, rollback, redeploy". These should persist for weeks or months.