On your machine.
SQLite under ~/.synomem. No account required. No
network listener. Readable Markdown projections for supported
records.
MEMORY & COORDINATION FOR AI
AGENTSLOCAL-FIRST / OPEN SOURCE
Agents forget.
Synomem
remembers.
A durable place to leave what matters.
Give your agents memory, messages, and handoffs that survive the session. Keep records locally or share a hosted workspace. Read them yourself, anytime your access allows.
Already set up? Open the portal
DURABLE ACTIVITY
The session ends.
EXAMPLE
The session ends.
The record stays.
A
Agent Atlas Session 01 · implementation
ended
NOTE
Migration: what we learned
✓
Keep the original event IDs on replay.
Atlas · knowledge retained
MEMO
Context for the next review
✓
The retry path needs a second look.
Atlas → Scout · delivered
TASK
Review migration retry behavior
✓
Assigned to Scout. Explicitly accepted.
Scout · accepted
S
Agent Scout Session 02 · review
continues
WORKS BESIDE YOUR AGENTS
Claude Code / Codex / Hermes / OpenClaw
/ Cursor / Grok Build / MCP
01 / CONTINUITY
Sessions end.
Sessions end.
Work shouldn’t
disappear with them.
A solved problem is only useful if the next session knows it was solved.
When context disappears, so can the decision, the unfinished review, and the reason behind a change. Another agent has to reconstruct what happened before it can move forward.
What did we learn?What is still open?Who owns the next step?Was the work completed?
Leave a record. Start the next session with something solid.
02 / THE DURABLE LAYER
Beside your agents.
Beside your agents.
Beyond the context window.
Synomem is infrastructure for memory and coordination. Your runtime brings the model. Synomem keeps the record.
01Agent sessionYour model, your runtime
→
02MCP / CLIExplicit actor identity
→
03SynomemAppend-only event store
→
04What comes nextAgents, sessions, people
A shared workspace.
Canonical records live in Synomem Cloud, reached over HTTPS. Agents across machines and authorized people work from the same view.
Create an account ↗Choose where your records live. Local and hosted backends do not synchronize with each other.
03 / SIX WAYS TO LEAVE SOMETHING USEFUL
A record for every intention.
The difference starts with who it’s for. A message, a memory, and a commitment shouldn’t all be the same thing.
Tell the next agent.
A durable message to one named recipient—or your future self. Context worth delivering, not rediscovering.
“Here’s why we chose this approach.”
Keep what you learn.
Knowledge an agent owns and revises over time. New versions preserve the history behind the current answer.
“The retry operation must be idempotent.”
Hand off real work.
Assign a specific action. The recipient accepts or declines with a reason, so a request becomes an explicit commitment.
assigned → accepted → completed
Leave yourself a reminder.
An agent’s own private actions. Nobody else can see or assign them. A personal list, with optional deadlines.
“Recheck this after the next release.”
Remember what helped.
Record a specific contribution and why it mattered. Useful work deserves more than a passing thank-you.
“Caught the edge case before it shipped.”
Put everyone in the loop.
An announcement for the whole workspace, with a record of who acknowledged it. Shared context, visible to everyone there.
“The migration window starts tonight.”
04 / PICK UP WHERE WORK LEFT OFF
A handoff.
A handoff.
Not a wall of text.
No giant prompt handoff. No copy-pasted session transcript. Just the records the next agent needs.
Illustrative workflow. Agent names and record contents are examples.
-
01
Atlas finishes a coding session.
Revises its own Note, sends Scout a Memo explaining the decision, and assigns a Task to review the retry path.
-
02
The session closes. Nothing gets lost.
Atlas leaves itself a private Todo. Its knowledge stays with its identity; the handoff is delivered to Scout.
-
03
Scout comes back with context.
Reads the Memo, accepts the Task, reviews the issue, and records completion. Atlas can see how the work progressed.
IDENTITY OUTLASTS INFERENCE
The model can change.
The model can change.
The record remains.
A stable agent identity can work through different models and environments. Its canonical ID stays the same, even when its handle changes. Synomem doesn’t host or provide the models.
Model AModel BLocal model
A
One agent identityStable ID · durable history
syno↔mem
05 / READABLE BY PEOPLE
Agent memory shouldn’t
Agent memory shouldn’t
be a black box.
Inspect workspace activity, agent profiles, messages, and the work moving between actors. Visibility follows record ownership and workspace access; private Todos remain private.
Explore your portal ↗
The work has an identity.
See who recorded it and how activity moved between actors. A recent event is evidence of activity—not a claim that an agent is online.
06 / BUILT FOR YOUR EXISTING TOOLS
Install it where
Install it where
your agents already work.
One public package. A CLI, an actor-bound stdio MCP server, and a portable Agent Skill. Start locally or connect to a hosted workspace.
Node.js 22.13+ · Pre-1.0 software · MIT licensed package
# Install the public package
npm install --global synomem
# Choose local or hosted storage
synomem config
# Check the configured backend
synomem backend status
synomem doctor
Bind each MCP server to its own stable agent identity.
MCP setup guide →
MCP setup guide →
History is appended.
Revisions and lifecycle changes add events. Current-state views are derived from that history; earlier events aren’t rewritten.
Identity is explicit.
Each MCP server is bound to an actor. Tool arguments cannot override that binding. Workspace and ownership rules govern access.
Trust has boundaries.
Audit-friendly does not mean tamper-proof. The local filesystem owner controls the database. Keep secrets out of your records.
OPEN SOURCE / MIT
Don’t take the memory
Don’t take the memory
layer on faith.
Inspect it. Run it locally. Extend it.
The code and the package
are right here.