TulimoaTulimoaDocs
Docs menu

Memory in the gateway

The memory tools every gateway connection gets for free: scopes and retention, automatic ID capture, protection rules and the dashboard view.

Last updated Sep 12, 2026

Every agent connected to the gateway gets a set of memory tools. The gateway answers them itself, so they work on every plan and never cost credits. Goals, decisions and IDs survive context compaction, session restarts and even a switch to a different client, because they live in the gateway, not in the agent's context window.

This is the gateway's own memory: it stores and recalls. Nightly consolidation, the morning briefing and the vault for personal data belong to Tulimoa Memory, which your agent can use in addition to the gateway.

The memory tools

All memory tools live in the tulimoa namespace. On connect the gateway teaches your agent a routine: recall at the start of a task, remember whenever something durable comes up, set_goal and checkpoint before multi-step work, get_pending after a reconnect, and recall once more before finishing.

tulimoa__recall(topic?, limit?)

Pulls back what was established earlier: pinned facts, IDs, decisions, open loops. topic filters by text, limit caps the items (1 to 25, default 8). Results are sorted by importance, newest first.

tulimoa__remember(text, key?, importance?, durable?, share_with_team?)write

Stores a fact as a short sentence. A later remember with the same key replaces it, importance weighs it (0 to 1, default 0.6), durable: true keeps it beyond the session, share_with_team: true shares it with your team if sharing is enabled for you.

tulimoa__set_goal(goal)write

Sets the one active goal for this session. Resurfaced if the agent's context gets compacted.

tulimoa__checkpoint(goal, plan_steps?, open_questions?)write

Saves a plan before a multi-step, multi-tool task so the agent can resume after compaction.

tulimoa__forget(id | key)write

Removes an item by its id or key, so recall no longer returns it. To delete it for good, use the Memories tab in the dashboard, see below.

tulimoa__get_pending(since?)

Replays what the agent may have missed while disconnected or after compaction, for example tool list changes.

tulimoa__recall_result(cache_key)

Fetches the full payload of an earlier tool result that was replaced by a short dedup pointer.

Tools marked write need a key or sign-in with the write scope. Reading memory works with every key.

Scopes and retention

ScopeHow it is setRetention
SessionThe default for remember, goals, checkpoints and captured IDs.Deleted after 14 days, on every plan.
Accountremember with durable: true.Kept until you delete it.
Teamremember with share_with_team: true, if sharing is enabled for you.Kept until deleted; visible to members who may read team memory.

Recall returns your account and team items plus the items of the current session. How team sharing is switched on per member is covered in Teams.

What happens automatically

  • A routine in the server instructions. On connect the agent receives a short instruction block that teaches it the memory routine, so it uses the tools without you prompting for it.
  • ID capture without an LLM. IDs are read only from the structured arguments and results of tool calls, never from free text: prefixed IDs and UUIDs, plus patterns known per connector. Up to 10 per call are kept in the session. That keeps prompt injection out of your memory.
  • A frugal context note. The gateway appends goal, pinned facts and IDs to a tool result only when memory changed, when the agent seems to lose track, or every 12 calls, in 140 to 300 tokens. It never spends tokens on every call.
  • Result dedup. When an identical call returns the same large result again, the agent gets a short pointer instead; tulimoa__recall_result fetches the full payload when it is actually needed.
  • Visible injections. Every automatic context note is listed under Usage with its reason: heartbeat, memory changed, disorientation or compaction anchor.

Protection rules

  • Credentials are never stored. If a text looks like a secret (a key, a token, a password), the gateway refuses it and asks the agent to store a reference instead.
  • Instruction-shaped text is quarantined. A memory that reads like a command is stored, but never handed back to an agent, neither in recall nor in the context note, so a poisoned memory cannot steer your agent.
  • Team writes are opt-in twice. A fact only reaches the team when the agent explicitly asks for it and your membership allows writing. IDs captured from tools never reach the team.

Memory in the dashboard

The Memories tab under Usage lists what your agents stored, without the automatically captured IDs: editable, deletable for good and downloadable as JSON. Under Settings you can export all your data (GDPR Art. 20) or erase your gateway memory completely (Art. 17); an erasure also reaches running sessions within about 30 seconds.