Agent Memory Governance

Architecture mémoire session

          
mermaid
          flowchart LR
  A[AGENTS.md] --> C[Session bootstrap]
  B[project_context.md] --> C
  D[latest-session.md] --> C
  C --> E[Execution session]
  E --> F[Session close update]
  F --> D
        

Fallback statique:

          
md
          ![Agent memory governance fallback](../../archive/fallback-agent-memory-governance.png)
        

Purpose

Keep agent context stable across sessions while minimizing token waste.

Source-of-truth layers

1. AGENTS.md (stable rules)

• Global operating rules and response style.
• Update rarely (only when process changes).

1. project_context.md (semi-stable project context)

• Architecture, stack, critical files, and validation commands.
• Update when architecture or workflows change.

1. documentation/sessions/history/latest-session.md (volatile session memory)

• Recent done, in-progress, next, and risks.
• Update at the end of each meaningful session.

Update cadence

• AGENTS.md: monthly or by explicit process decision.
• project_context.md: when runtime topology or conventions change.
• documentation/sessions/history/latest-session.md: every session close.

Session protocol

Start:

1. Run npm run session:bootstrap.
2. Run npm run session:budget.
3. Confirm scope and risks.

Close:

1. Run npm run session:close -- --done "..." --next "..." --risk "...".
2. Keep entries short and factual.

Anti-drift rules

• Do not duplicate stable rules in latest-session.md.
• Do not store volatile TODO lists in AGENTS.md.
• Keep each section concise and deduplicated.
• Keep documentation/sessions/history/latest-session.md below line budget (default cap: 140 lines, 8 items/section).