SSyncDev

Recall MCP Server

Agent memory that keeps every correction with its history, plus saved working state so a crashed run can resume.

Local server

What is the Recall MCP Server?

Most agent memory tools overwrite a fact when it changes. Recall treats a correction as a new record. Ask what the agent believes about your editor and it can answer now, as of a past date, and with what the answer replaced. The Recall MCP server is one face of a Go library and a recall binary that also serves HTTP, alongside a Python client.

Memory: facts with a past

A memory is a subject, a predicate and a typed value. Predicates sit in a registry that sets their type, whether they hold one value or many, and which ones contradict each other. Forgetting records a withdrawal rather than deleting anything. Starter registries exist for coding, assistant, support, sales and voice agents. If you would rather not define predicates, an open vocabulary lets first use define them. With an extraction model configured, the MCP tools take plain text: remember, recall and forget. Without one, recall mcp -typed serves typed tools instead.

State: resuming after a crash

The second half, currently beta, records where an agent is: goal, plan, progress, verbatim turns, pointers to branches or processes, and each world-touching tool call, logged before it runs. A new process resumes from those records within a token budget, and any action left unfinished shows up as unconfirmed rather than being repeated. A lease ensures one process acts for an agent at a time. The MCP tools here include agent_resume, update_working_state, record_turn and begin_action.

Installing and choosing storage

Install via pip install polign-recall, Homebrew, or go install, then register recall mcp -write with your client. With nothing configured, memory sits in a local database under ~/.local/share/recall, shared by every agent on the machine, and no model or API key is required to start. To use infrastructure you already run, set RECALL_BACKEND and RECALL_BACKEND_URL:

Backend Keyword search State API
local (default) yes yes
polign_db yes yes
PostgreSQL + pgvector yes yes
MongoDB yes yes
Qdrant no no

For many agents, recall serve holds the database credentials and hands each agent a namespaced key.

Points to note

Release builds send an anonymous usage heartbeat with version, platform and operation counts; RECALL_TELEMETRY=0 or DO_NOT_TRACK=1 disables it. The State API is still beta. Browse our other integrations or see comparable tools in that category.

Recorded tools

Check the linked publisher documentation and your connected client's tool list. Names and availability can vary with the server version and account permissions.

ToolWhat it does
rememberStores a fact, as plain text when an extraction model is configured.
recallAnswers what the agent currently believes, with what each answer replaced.
forgetRecords a withdrawal of a fact without deleting its history.
agent_resumeResumes an agent from its recorded state within a token budget, taking a lease.
update_working_stateUpdates the agent's goal, plan, progress and related working state.
record_turnStores a conversation turn verbatim.
begin_actionRecords a world-touching tool call before it runs so a crash leaves it visible.

How to install the Recall MCP Server

{
  "mcpServers": {
    "recall": {
      "command": "recall",
      "args": [
        "mcp",
        "-write"
      ]
    }
  }
}

Install first with pip install polign-recall, brew install polign/tap/recall, or go install.

Configuration

VariableDescriptionRequired
RECALL_BACKENDStorage backend: local (default), polign, pgvector, mongodb or qdrant.Optional
RECALL_BACKEND_URLDirectory or connection URL for the chosen backend.Optional
RECALL_URLAddress of a recall serve instance, used by recall connect.Optional
RECALL_API_KEYRecall key for a recall serve instance, used by recall connect.Optional
RECALL_TELEMETRYSet to 0 to turn off the anonymous usage heartbeat.Optional

Frequently asked questions

Not to start. With nothing configured, Recall runs a local database on first use. An extraction model is optional and lets the MCP tools accept plain text instead of typed predicates.