Files
fidelity-ai-workspace/docs/memory-model.md

2.1 KiB

Memory Model

AI Workspace separates memory by purpose so that human-readable knowledge, raw evidence, agent behavior, and derived indexes do not collapse into one opaque store.

Memory Layers

Layer Purpose Examples Canonical?
Canonical project memory Durable project facts for humans and AI current work, work items, people, decisions, daily notes yes
Raw evidence Captured data before curation Mattermost mirror, photos, archives no
Agent operating memory Rules for agent behavior and workflows promotion rules, communication style, verification rules yes for agent behavior
Derived index Fast retrieval over canonical memory .aiw/indexes/<profile>/ no
External agent memory Optional cross-agent recall mem9, tool auto-memory no for project truth

Canonical Markdown

The project knowledge vault should be readable without any AI tool:

project-knowledge/
  00-start/
  01-current/
  02-work-items/
  03-context/
  04-people/
  05-decisions/
  06-daily/
  07-maps/
  08-bases/
  09-templates/

For future multi-profile setups, this same structure should live under a profile-specific workspace path.

Evidence Promotion

Connectors write evidence. They do not decide what becomes memory.

Promotion flow:

inbox evidence → verified fact → smallest correct Markdown file

Do not promote:

  • secrets;
  • sync status;
  • generic chatter;
  • unverified guesses;
  • raw transcripts without curation.

Local Index

The local RAG index is derived from canonical Markdown. It helps AI clients find relevant snippets quickly, but it is not the source of truth.

If index output conflicts with Markdown, Markdown wins.

mem9 And Similar Memory Systems

mem9 can be useful as an optional cross-agent recall layer for preferences, reusable workflow habits, and non-sensitive operational memory. It should not replace the project knowledge vault.

Recommended stance:

project-knowledge/ wins over mem9, vector stores, and chat memory.

For sensitive or corporate projects, avoid cloud memory ingestion unless the data policy is explicit and approved.