Local Agent Memory v1

Build or refine a reliable file-based memory system for an agent.

Core workflow

1. 1. Create or inspect these layers:

1. 2. Keep MEMORY.md lightweight and routing-oriented.
2. Put stable facts in semantic files.
3. Put repeatable methods in procedural files.
4. Treat memory as a hint/index layer, not unquestionable truth.
5. Re-verify current facts before taking real actions based on remembered information.
6. Write destination files first, then update MEMORY.md only if the change deserves long-term indexing.

Decision rules

Use daily memory for

• - new events
• one-off attempts
• temporary troubleshooting detail
• anything not yet proven reusable

Use semantic memory for

• - stable user preferences
• durable environment facts
• platform constraints
• lasting architecture or governance decisions

Use procedural memory for

• - repeatable workflows
• checklists
• maintenance routines
• methods likely to be reused across sessions

Maintenance pattern

Run a lightweight dream/consolidation pass when memory starts to sprawl:

• - read INLINECODE6
• read recent daily logs
• identify repeated facts or workflows
• extract stable facts into semantic memory
• extract repeatable methods into procedural memory
• prune low-value or duplicated summary lines from INLINECODE7

Run a deeper pass for large daily logs or when the topic tree needs restructuring.

Guardrails

• - Do not let MEMORY.md become a diary.
• Do not promote everything that looks interesting.
• Do not rely on stale remembered facts for real actions.
• Do not mix memory maintenance with unrelated code changes unless the user asked for both.
• Prefer a few clear topic files over many overlapping files.

References

Read these only as needed:

• - references/architecture.md for the memory model and core disciplines
• INLINECODE10 for minimum structure and topic layout
• INLINECODE11 for governance and consolidation rules