m2ai-portfolio Agent Memory Architecture Decision Matrix
Guides a builder from raw constraints to a concrete memory architecture choice, then produces an ADR they can commit to their repo.
When to Invoke
Trigger on: "how should I store agent memory", "which memory backend", "Claude memory vs Mem0 vs Zep", "design memory layer", "memory architecture ADR", "persistent agent state".
Decision Dimensions
Ask the user (one at a time, inline):
- Portability requirement — must the memory survive platform migration? (yes/nice-to-have/no)
- Latency budget — is sub-100ms recall mandatory, or is 1-2s acceptable?
- Memory volume — rough count of distinct memories after 6 months (hundreds / thousands / millions)
- Recall style — keyword search / semantic search / structured queries / mixed
- Team size maintaining it — solo builder / small team / dedicated platform team
- Budget — hosted spend tolerance per month
- Compliance — any data residency, PII, or audit requirements
Backend Candidates
Score each on a 1-5 scale per dimension:
| Backend | Portability | Latency | Scale | Maintenance | Cost | Compliance |
|---|---|---|---|---|---|---|
| Provider-hosted (Claude/GPT/Gemini memory) | 1 | 5 | 5 | 5 | 4 | 2 |
| Mem0 (hosted) | 3 | 4 | 4 | 4 | 3 | 3 |
| Zep (hosted or self-host) | 4 | 4 | 4 | 3 | 3 | 4 |
| SQLite + embedding column | 5 | 3 | 3 | 3 | 5 | 5 |
| Postgres + pgvector | 5 | 4 | 5 | 2 | 3 | 5 |
| Custom JSON files + FTS | 5 | 2 | 2 | 2 | 5 | 5 |
Weighting
Multiply each backend's scores by the user's stated priorities, then rank. Tie-break by portability (the most-underweighted axis per Nate's 2026-04-08 thesis).
Output — ADR Template
# ADR: Agent Memory Architecture
**Date**: {date}
**Status**: Proposed
## Context
{1-paragraph summary of the agent, its memory use case, and the constraints the user stated}
## Decision
We will use **{chosen backend}** for persistent memory because:
- {top reason tied to the user's highest-weighted dimension}
- {portability stance — explicit}
- {latency stance — explicit}
## Alternatives Considered
| Backend | Score | Why Not |
|---------|-------|---------|
| ... | ... | ... |
## Portability Plan
How memories are exported and re-imported if we migrate. If the chosen backend is provider-hosted and non-portable, **this section must call that out as a known risk and specify a snapshot cadence.**
## Rollback
What replaces this if the backend disappears, is acquired, or changes terms.
Verification
- All 7 dimensions were asked (or had sensible defaults applied with a note)
- The chosen backend has the highest weighted score
- Portability Plan is non-empty even if the answer is "we accept the lock-in"
- ADR saved to a repo path the user specifies, or to
/tmp/adr-agent-memory-{date}.md
Source
Nate's Newsletter — "512,000 Lines of Leaked Code Reveal the Lock-In Strategy Coming for Your AI Stack" (2026-04-08). The thesis that behavioral/memory context is the most-underrated lock-in axis shaped the tie-break rule above.