Skip to main content
Slowave exposes its memory layer to AI coding agents through five MCP tools that together form a deterministic cognitive cycle. Every task begins with slowave_activate, proceeds through optional slowave_remember and slowave_recall calls during the work, records assessments with slowave_feedback, and closes with slowave_commit. The tools are served by a local HTTP daemon at http://127.0.0.1:8766/mcp — nothing leaves your machine.

The five tools

Cognitive cycle order

Each task follows the same five-step sequence. The steps are numbered to reflect their intended order; not every step is mandatory on every task, but the order must never be reversed.
1

activate

Call slowave_activate once at the very start of the task. It opens a server-side session, runs spreading activation over scoped memory, and returns a compact working-memory set plus a retrieval_id and session_id you will use in all subsequent calls.
2

remember

Call slowave_remember any time during the task to persist a durable, standalone claim — a fact, decision, preference, constraint, lesson, or warning — that should survive beyond the current session. Ephemeral state does not belong here.
3

recall

Call slowave_recall when the task evolves and you need specific historical context that activation did not surface. Each recall returns a fresh retrieval_id that must be covered by feedback before committing.
4

feedback

Call slowave_feedback after using (or deciding not to use) each set of retrieved memories and procedures. Assessments are append-only. Every target exposed by activation or recall must receive feedback before the session can be committed.
5

commit

Call slowave_commit at the end of every task. It validates that all feedback is complete, records the outcome and verification, optionally captures a reusable procedure, and triggers offline memory consolidation.

MCP server and client configuration

The Slowave HTTP MCP daemon runs locally and accepts connections only from 127.0.0.1. Start or check its status with the CLI:
Add the following block to your MCP client configuration to connect:
The daemon port and bind host can be overridden with environment variables:
Run slowave setup --client <name> to have Slowave write the MCP configuration block automatically for your client. Supported clients include claude-code, claude-desktop, cline, cursor, windsurf, opencode, and codex.

Scope format

Every scoped tool call requires a scope parameter in kind:id form. Both the kind and id parts must be nonblank. Scopes create strict memory isolation — retrieval for one scope never returns memories from another unless a schema has generalized through cross-scope evidence over time.
Scope strings are case-sensitive and treat - and _ as distinct characters. project:my-repo and project:my_repo are two separate, fully isolated memory stores. If activation returns a scope_fragmentation warning, check whether you have inadvertently created a variant scope.

Session lifecycle

A session is opened implicitly by slowave_activate and closed explicitly by slowave_commit. The session_id returned by activate must be passed to remember, recall, and commit. You never call a separate session-start command from the MCP surface. If slowave_commit is not called, the idle-session reaper closes the session automatically after SLOWAVE_SESSION_IDLE_TIMEOUT seconds (default 3600). Reaper-closed sessions do not enforce feedback completeness.

continuity_id and cross-conversation correlation

continuity_id is an opaque token Slowave issues to correlate multiple activations that belong to the same client conversation thread. The rules are simple:
  • Omit it on the very first slowave_activate call in a new conversation.
  • Retain and resend the continuity_id value from the activate response on every subsequent activation within the same conversation.
  • Never invent a continuity_id or reuse one from a different conversation.
Continuity does not relax the scope boundary. It only tells Slowave that two activations are part of the same reasoning thread, which may expand context reinstatement on the first activation of a new task.

MCP tools vs CLI lifecycle

The CLI uses an older five-step sequence that includes reinforce instead of feedback. Do not use CLI examples as an agent integration contract. The MCP tools — activate → remember → recall → feedback → commit — are the only supported path for agent integrations. slowave reinforce is a CLI compatibility command and is not an MCP tool.