1. The Problem Memory Base Solves
Standard AI chats are amnesiac. Every new chat thread forces you to repeat your role, preferred tech stack, and project rules.
Scoped strictly to one conversation thread. Dies when starting a new chat.
Scoped strictly to chats inside a specific folder (e.g. "In this folder, use TypeScript").
Saved facts can be included in relevant chats while active. They remain in browser storage until removed or site data is cleared; optional backup and sync depend on your settings.
2. Opening & Navigating Memory Manager
Click Memory (Brain icon) in the left sidebar to open the full-screen Memory Base Manager.
- Preferences: Preferred language, framework, styling, and coding conventions.
- Personal: Role, location, timezone, and communication style preferences.
- Projects: Long-term facts and architectural constraints for specific active projects.
- General: Universal facts, domain knowledge, or custom workflow rules.
3. AI Memory Suggestions — Review Before Saving
When memory suggestions are enabled, extraction is attempted only for prompts containing one of the app's memory cues (such as “remember” or “my preference”). The configured completion route receives the eligible prompt and assistant response to suggest facts; it does not run after every response. Suggestions appear above the chat composer and are saved only if you choose Keep within 60 seconds. No more than two AI-created items can be proposed in one conversation.
✅ Possible Memory Suggestions:
- "I live in Allentown, PA" → Proposed for the Personal category; requires Keep.
- "I always write TypeScript with strict null checks" → Proposed for Preferences; requires Keep.
❌ Ignored Short-Term Context:
The extraction instructions ask the model to ignore temporary details, code snippets, and attachments. Review every proposal before it is retained; discard anything you do not want saved.
4. Context-Aware Relevance Filtering
To prevent flooding system prompt tokens, Accelerated Logic AI performs keyword relevance matching before each turn. Only memories matching your current prompt context (or category) are injected into the model payload. Asking "What do you remember about me?" automatically injects all active memories.
5. Manual Fact Etching & Presets
Click Manual Entry to etch specific facts. On fresh installations, select from preset templates (Default Stack, Aesthetics, Developer Role, Architecture) for instant 1-click memory initialization.
6. Managing Memory Cards
Mute/unmute individual memories without deleting them.
Edit phrase wording or update category classification.
Purge all long-term facts with a single confirmed action.
7. Memory Settings & Deletion Permissions
In Settings → Fact Memory, toggle AI Can Delete Memories. When enabled, saying "Forget that I live in NY" allows the model to output a <delete_memory id="..."> tag to remove outdated facts safely.
8. Memory Best Practices
Do Save: Standard stack preferences, coding conventions, persona roles, WCAG accessibility rules, and long-term project requirements.
Do Not Save: Temporary bug details, raw file contents, secrets or API keys, or conflicting preferences (e.g. both light and dark mode).
9. Cloud Sync & Workspace Export
When Google Drive backup is enabled with Memories toggled ON, facts sync automatically to my_ai_memories.json. You can also export or import standalone JSON files in Settings → Export Workspace.
10. Memory Troubleshooting
Memory Not Applying? Ensure the Eye icon is active (not muted) and verify the memory text includes explicit keywords (e.g. "TypeScript") for relevance matching.
Conflicting Memory? Search for old preferences and delete outdated entries to prevent contradicting instructions.
11. Writing Memories That Actually Change Model Behavior
A memory is injected verbatim into the system prompt — so write it like a spec, not like a diary entry. The difference between a helpful and a useless memory:
- "I like good code" — every preference cancels out
- "Working on a startup" — no constraint, no stack
- "Be concise" — unmeasurable; the model cannot tell if it succeeded
- "TypeScript strict mode; no
any; named exports only; Tailwind for styling" - "Project Atlas: Next.js App Router, Postgres via Drizzle, deployed on Vercel — avoid Pages Router patterns"
- "Answer with a short rationale, then code, then a 2-line summary; never restating my question"
Template worth memorizing: [Scope] [Entity] [Hard rule] [Why] — e.g. "For Atlas, always use Drizzle query builder over raw SQL — the team standardized after two injection near-misses." Memories with a reason survive contradictions better because the model can weigh them against context.
Keyword hygiene: memories are relevance-matched on terms. A memory about "the Atlas repo" will not fire on a prompt that never mentions Atlas — include the project name, the file type, or the domain noun it should trigger on.
12. Memory Lifecycle — Auditing, De-conflicting, Pruning
Memory bases rot silently: three months of accumulated facts contain contradictions and dead projects. A quarterly audit takes ten minutes:
- Export first. Settings → Export Workspace → Memories JSON. Now you can experiment destructively without fear.
- Read for staleness. Delete anything pinned to finished projects, expired deadlines, or abandoned stacks ("Remix v2 patterns" when you now ship Vite).
- Hunt contradictions. Search each pair of opposing terms ("light mode"/"dark mode", "jest"/"vitest", "tabs"/"spaces") and keep the winner.
- Check relevance scope. Rewrite memories that are too broad ("be careful with SQL") into scoped, keyword-rich forms ("Atlas uses Drizzle — avoid raw SQL in src/db/").
- Watch the size. Every active memory is injected context. A lean base of 30 surgical facts beats a 300-fact attic where real preferences drown.
13. Memory vs. Folder Memory vs. Chat History vs. Skills
| Mechanism | Scope | Lifetime | Best for |
|---|---|---|---|
| Memory Base | Global, relevance-matched | Until you delete it (or wipe) | Cross-project facts: stack, style, role, conventions |
| Folder Memory | One folder's chats | Until folder memory is edited/cleared | Per-project rules: "Atlas always uses Drizzle" |
| Chat History | One thread | Thread lifetime | Turn-by-turn context; dies with the thread |
| Skills | Global, opt-in per use | Until deleted from Skills Manager | Reusable procedures & expertise, toggled per task |
| Pinned Files | One chat | Until unpinned | Files every prompt in this thread needs |
Rule of thumb: memories answer "who are you / how do you work", folder memory answers "what are we building here", history answers "what just happened", skills answer "how do we do this task".
14. Privacy Review — What Your Memory Base Actually Contains
Because memories are injected into prompts, every memory is a disclosure to whichever provider you chat with. Do this review once:
- Open the Memory Manager and read every entry. Yes, all of them — you are auditing your own prompt annex.
- Remove or rewrite anything you would not paste to a stranger: home address, salary, health details, another person's name tied to real decisions, anything under NDA from an employer.
- Check the sync path. If Google Drive backup is enabled for Memories (Part 17), these facts also rest in Google's AppData vault under your account — appropriate for most, worth knowing for all.
- Two export doors exist: the full workspace snapshot (Part 17) and the per-slice memories export. The per-slice file is plain JSON — treat it exactly like the database it is.
Never do: "remember my API key is sk-..." — memories ship in prompts. Keys belong in the credential store (Part 22), never in memory text.