Memory Architecture Spec Template
A one-page specification of an agent memory system: types, storage, retrieval, decay, and privacy, decided before the first write. Copy it below, fill the bracketed fields, delete what you don't need.
# Memory architecture: [system name]
Date: [date] · Owner: [name]
## Memory types in scope
- [ ] Episodic (what happened): [store / not stored]
- [ ] Semantic (facts): [store / not stored]
- [ ] Procedural (how-to): [skills directory / other]
- [ ] Working (task state): [per-task files, deleted on completion]
## Storage and retrieval
| Tier | Contents | Loaded | Store |
|------|----------|--------|-------|
| Hot | [index, digest] | every session | [flat file] |
| Warm | [fact bodies] | on relevance | [files / db] |
| Cold | [archives, logs] | explicit query | [store] |
## Write rules
- Written by: [agent / hooks / humans]
- Never written: credentials, personal data, [additions]
- Every fact carries: date, source
## Decay and staleness
- Facts are treated as claims to re-verify on recall: yes
- Consolidation pass cadence: [weekly / monthly]
- Supersession over deletion for decisions: yes
## Privacy and security
- Access control: [who/what reads each tier]
- Redaction happens at: write time
## Evaluation
- Repeated-mistake count tracked via: [method]
- Staleness audit: [sample N facts per month, measure still-true rate]
Field notes
| Field | Why it matters |
|---|---|
| Write rules | The write policy is the security boundary. Deciding what may never be written, before the first write, is the cheapest privacy control you will ever implement. |
| Decay and staleness | The section most systems skip and the reason most memory systems rot. If you fill in only one section honestly, make it this one. |