Knowledge & Memory · Concept
Agent Memory Trust Contract: Remember, Cite, Forget
Source status: derived from Vox's X article, "A Framework for Agent Memory: Remember, Cite, Forget" (2026-05-22), reviewed from the live X article page. Treat it as a practitioner framework, not a formal standard. Agent memory is reliable only when it does three jobs at once: Remember what should persist. Cite the source and authority level of what was remembered. Forget or demote memories that are stale, superseded,…
wiki/wiki/concepts/agent-memory-remember-cite-forget.mdAnswer
Source status: derived from Vox's X article, "A Framework for Agent Memory: Remember, Cite, Forget" (2026-05-22), reviewed from the live X article page. Treat it as a practitioner framework, not a formal standard. Agent memory is reliable only when it does three jobs at once: Remember what should persist. Cite the source and authority level of what was remembered. Forget or demote memories that are stale, superseded,…
Auto-generated neutral summary from the source page — needs human review before trusted use.
Evidence & Source Cards
https://x.com/Voxyz_ai/status/2057862044783628508external/unverifiedSource Excerpt
Source status: derived from Vox's X article, "A Framework for Agent Memory: Remember, Cite, Forget" (2026-05-22), reviewed from the live X article page. Treat it as a practitioner framework, not a formal standard.
Core Thesis
Agent memory is reliable only when it does three jobs at once:
- Remember what should persist.
- Cite the source and authority level of what was remembered.
- Forget or demote memories that are stale, superseded, or outside their authority.
The useful phrase is: remember by layer, cite by source, forget by expiry. A memory without a layer, provenance, and expiry rule should remain background context rather than control an agent's decision.
Memory Layers
| Layer | Purpose | Typical Lifespan | Failure Mode |
|---|---|---|---|
| Hot session | Current task working context: recent user message, tool outputs, intermediate conclusions | One task/session | Context compression drops an important instruction, then the agent acts as if it was never said |
| Day-state | Today's operating whiteboard: what has been done, what is next, current sequence | Hours to one day | A cron agent continues executing an old morning plan after a later direct instruction changed priority |
| Project memory | Long-running lessons, settled decisions, repeated pitfalls | Days to months | Old preferences keep shaping behavior after newer preferences superseded them |
| Retrieval/index | Search layer that surfaces candidates from vector, graph, or text search | Rebuild-dependent | An old plan retrieved by similarity is mistaken for the current plan |
| Canonical policy | Manually maintained rule layer: AGENTS.md, SOUL.md, policy files, product boundaries | Long-lived, versioned | A summary or stale memory overrides a real policy file |
| Direct instruction | Immediate user command for the active task | Current task, until superseded | A paraphrased summary is treated as authorization even though the original confirmation is not traceable |
Authority Order
When memory sources conflict, use an explicit authority order. For Clark Farming Company agent work, a safe default is:
- Original direct instruction from the user for the active task.
- Canonical policy committed in doctrine, project policy, or governance files.
- Most recent project decision with traceable source.
- Long-term memory with attribution and timestamp.
- Retrieval summary that points to underlying source material.
- Compressed summary with no direct citation.
Higher levels overrule lower levels. A summary saying "the user seemed to authorize this" does not override a policy requiring explicit authorization.
Expiry Mechanisms
| Mechanism | Description | Best For |
|---|---|---|
| Hard expiry | Facts include valid_from / valid_until; a new fact marks the old one expired while retaining history | Policy, pricing, regulations, credentials, model/provider capabilities |
| Bitemporal facts | Track when the fact was true in the world and when the system learned it changed | User/project facts that can be corrected later; audit-sensitive knowledge |
| Soft decay | Retrieval ranking demotes old or unused memories without deleting them | Preferences, habits, non-critical context |
Do not design storage first and expiry later. The expiry rule is part of the memory schema.
Decision Gate Before Memory Can Influence Action
Before an agent lets a memory affect a decision, answer three questions:
- What level of decision can this affect? Hint only, citable evidence, or final authority?
- Where did it come from? Which message, file, page, or source created it?
- Is it still valid? What newer information would make it step aside?
If the agent cannot answer all three, the memory may be retrieved for context but should not control the outcome.
Application to Hermes / Clark Farming Company
SOUL.md, skill files, and wiki policy pages are canonical policy, not ordinary memories.- Persistent memory entries should stay compact because they can become decision-shaping context in future sessions.
- Session transcripts and
state.dbare evidence reservoirs; summaries should be treated as search aids, not as authority. - Wiki pages are durable project memory. They should cite sources, carry status/confidence, and be updated when superseded.
- Cron jobs need day-state discipline: today's output and next action should not silently inherit yesterday's plan unless the job explicitly says so.
Governance Assessment
This framework matches the company's standing rule: evidence before claims. The practical value is not the taxonomy itself; it is the trust contract it forces every memory to carry. An agent that remembers without citation will eventually hallucinate authority. An agent that cites without expiry will eventually act on stale truth. An agent that forgets without preserving source history will eventually lose accountability.
References
- Vox (@Voxyz_ai), "A Framework for Agent Memory: Remember, Cite, Forget," X article/post, 2026-05-22: https://x.com/Voxyz_ai/status/2057862044783628508
- Related internal entries: concepts/ai-memory-and-context-management-index, concepts/episodic-memory-for-ai-agents, concepts/semantic-memory-for-ai-agents, Hermes Agent SQLite Memory Infrastructure.
Relationships
Outbound links
- AI Agentscorpus
- AI Memory and Context Managementcorpus
- Episodic Memory for AI Agentscorpus
- Hermes Agent SQLite Memory Infrastructurecorpus
- Local-First Hermes Runtime Architecturecorpus
- Semantic Memory for AI Agentscorpus
Referenced by
- AI Memory and Context Managementbacklink
- Clark Farming Company Canonical Namingbacklink
- Wiki Frontmatter Maintenancebacklink