Agent Memory overview

What is Agent Memory?

Agent Memory lets an AI assistant remember facts about you across sessions instead of re-learning them every conversation — preferences, decisions, and project context that would otherwise be lost when the chat ends. Facts are stored as short, atomic statements (for example, "Prefers concise answers" or "This repo's staging DB is on Supabase"), each with its own provenance, salience, and recall ranking.

Available on Starter and above; requires a signed-in Gryz account.

Personal vs. project scope

Every memory is stored under one of two scopes (team members also get a third, private one, described below):

• Personal — facts about you that apply everywhere you use the assistant (preferences, working style, standing decisions).
• Project — facts scoped to a specific codebase or workspace, keyed by a derived project_key (from the git remote, repo path, or a slug you provide). Project memories only surface when the assistant is working in that same project.

agent_onboarding derives the right project_key for the current workspace and returns an idempotent rule block to place in your agent's config (AGENTS.md, CLAUDE.md, or a Cursor rules file) after explicit approval — see placement guidance for exact semantic targets. The old setup_memory name remains only as a compatibility alias.

Private memory for a team member

If you use Agent Team, each team member also has its own private memory, saved with scope: "agent".

• A member can read and write only its own private memory. Other members cannot see it, and cannot tell that it exists.
• A member never sees your personal memory.
• You can read, add and delete a member's private memory from its page in the dashboard.
• It counts toward your total memory limit. There is no separate limit per member.

Work with Agent Memory from an AI assistant

After you connect Gryz through MCP, an assistant has five Agent Memory tools:

• remember stores an atomic fact (scope, body, optional project_key, salience).
• recall retrieves scoped memories ranked by relevance, salience, and recency — returned inside an untrusted-data delimiter, since memory bodies can contain agent- or human-authored text.
• list_memories lists memories with a provenance summary, for browsing rather than ranked retrieval.
• forget tombstones a memory by id — retractions are tracked, not silently deleted.
• agent_onboarding derives project_key and proposes the rule block above; it never writes to disk itself.

See the MCP tool reference for exact inputs, safety annotations, and rate limits.

Plan limits

Every write (remember, POST /api/memory/, and JSONL import) is gated by two independent quotas — hitting either returns an error rather than silently dropping data:

Personal-scope memories share the same total-count quota as project-scope ones — it is not a separate allowance per scope. See the Memory API reference for the full table and REST equivalents.

Manage memories in the dashboard

In My Docs → Agent Memory, you can browse, filter by kind, sort, and inspect the full provenance of any stored fact, or delete one directly. Memory Insights surfaces health signals like near-duplicate flags and recall frequency across your stored facts.

Terms and further reading

• Agent Memory terms — glossary of scope, salience, provenance, and recall-ranking terms used throughout this section.
• Memory API reference — REST equivalents of the MCP tools, for the dashboard's same-origin fallback.