Project-Scoped Memory for AI Chat: A Walkthrough of aiFetchly Workspace Memory

• Three kinds of memory, kept apart
• What it remembers
• How scoping works
• Creating a memory by hand
• Reviewing and pruning
• Workspace auto-summary
• Settings you control
• Best practices

Every new AI chat starts from zero. You re-explain the product, the naming rules, the “don’t touch the staging database” warning, and the command that actually runs the tests. Then you open a second project and the assistant happily mixes up both.

aiFetchly, a desktop AI business agent, handles this with Workspace Memory: project-scoped memory for AI Chat that stays inside one approved workspace and never leaks into another. This guide walks through what it stores, how scoping works, and how to keep it clean.

§Three kinds of memory, kept apart

aiFetchly separates:

• Global user memory — facts about you that apply everywhere.
• Conversation memory — context inside a single chat.
• Workspace memory — durable, project-specific knowledge tied to one project folder, campaign workspace, or repository.

Workspace memory is the one that answers “what should any chat in this project already know?”

§What it remembers

Each memory has a type:

Just as important is what it should not hold: secrets, API keys, cookies, private scraped lead data, full transcripts, bulky tool output, raw file contents, or temporary task progress. Memory is for durable guidance, not a dump.

§How scoping works

Workspace memory only exists when the conversation has an approved workspace. aiFetchly resolves that folder to a stable identity:

• Inside a Git repository → memory is scoped to the repository root.
• No Git root → memory is scoped to the selected real path.

So two chats on the same approved workspace share memory, while memories from one workspace are never injected into another. No approved workspace means no workspace memory at all, which is a sensible default.

§Creating a memory by hand

1. Open AI Chat and choose or approve a workspace above the composer.
2. Click Memory on the workspace badge (the badge also shows the active count).
3. Click Create memory, pick a Type, write a short Title and concise Content.
4. Set Confidence (0–100; manual memories default high, lower it for shaky notes).
5. Save.

Example entries for a marketing workspace:

• Convention — “Brand name is always written aiFetchly, lowercase a.”
• Warning — “Never send campaign emails to the EU list without the double opt-in flag.”
• Workflow — “Run npm run build && npm run serve to preview docs locally.”
• Reference — “Pricing decisions live in docs/pricing-2026.md.”

§Reviewing and pruning

The dialog lists active memories with type, title, content, source (Manual, Chat, Agent task, or Auto-dream), updated time, last-used time and confidence. You can search by title or content, and toggle Show archived to see archived or contradicted entries.

Row actions:

• Edit — change type, title, content, confidence, or status (active / archived / contradicted).
• Archive — hide it without deleting.
• Delete — remove it permanently after confirmation.

The docs’ advice is good: archive outdated memories instead of leaving contradictory active guidance around.

§Workspace auto-summary

Writing memories by hand gets old. Workspace auto-summary (called “auto-dream” internally) consolidates useful project-specific information from your conversations and agent tasks into decisions, workflows, references and warnings. Click RUN AUTO SUMMARY for a manual run; the dialog shows the last run time. Everything it creates stays inspectable and editable.

§Settings you control

In System Settings → AI preferences:

• Workspace Memory Injection — inject relevant workspace memories into AI Chat context.
• Workspace Auto-Summary — background consolidation.
• Manual Workspace Memory — allow hand edits (turning it off doesn’t delete anything).

And one rule that matters: the current message wins. If a stored memory conflicts with what you ask right now, your current instruction takes priority.

§Best practices

• Store decisions and conventions you’d otherwise repeat in every chat.
• Store project commands that aren’t obvious from package scripts.
• Store warnings for compliance, safety, environment issues or test traps.
• Keep each memory short enough to be useful in prompt context.
• Link to file paths or docs instead of pasting file contents.

Full docs: https://docs.aifetchly.com/docs/ai-outreach/workspace-memory/

aiFetchly: https://www.aifetchly.com

If you try it, start with three memories: one convention, one workflow, one warning.