Agents write to the ledger
Turns, memory, decisions, validations — via MCP tools your agents call naturally. Append-only, hash-chained, enforced at the database layer.
Open-source MCP server · runs local
Your agents stop forgetting. Every handoff, verified.
Agent memory, with receipts. Continuity gives Claude, Codex, and MCP-compatible agents shared task state across sessions and models — derived from a tamper-evident ledger, so every handoff is generated from what actually happened, not what a summary remembered.
Test agent memory yourself
$ pip install continuity-mcp
Create a task, close the session, resume from the receipt — see exactly what the next agent remembers and why. Live, no signup.
M01 / DRIFT
Agents already write handoffs to themselves. Then the handoffs lie.
Compaction summaries, CLAUDE.md notes, a hand-tended HANDOFF.md — every one is a self-report. It keeps the decision that got reversed, drops the constraint that mattered, and when two sessions disagree, the last write wins in silence.
Summary-based resume
“Refactor going well.”
constraint: keep the v1 API frozen
next: rewrite in tRPC ← reversed 2 sessions ago
conflict with parallel session: unrecorded
Ledger-based resume
state: API migrated; tests green (seq 17)
constraint: v1 API frozen (seq 9, still active)
rejected: tRPC rewrite (seq 14, superseded seq 16)
conflict: surfaced at seq 20, resolved seq 21
Resume tools summarize.
Continuity replays what actually happened.
M02 / PROCESS
State is never stored. It's replayed.
Context is a projection
Task state, project memory, and conflicts are folded from events — the stale path and the reversal are structurally impossible to forget.
Any model picks it up
Out of Claude credits mid-task? Open Codex, compile the task context, keep working. Same task, same memory, zero rebriefing.
The receipt travels
A verified HANDOFF.md with integrity status, ledger tip hash, and traceable source rows. Commit it, send it, attach it to an invoice.
M03 / RECEIPT
The receipt is computed, not asserted.
Continuity does not ask the next agent to trust a summary. Every export replays the whole ledger and re-checks every hash. If the chain is valid, the receipt says verified — with the tip hash to prove it.
Alter one event and the same exporter reports integrity_failed with the broken sequence. Tamper the ledger and the receipt says so.
Billing a client, reviewing a decision, or reconstructing a dispute — the record is already there: which event, which agent, which hash.
HANDOFF.md · export
VERIFIED
task: checkout-refactor status: Ready state: API migrated; tests green next: ship behind feature flag rejected:rewrite in tRPC (seq 14)decided by: codex · session-a → claude · session-b integrity: provenance_status: verified ledger_tip_hash: 9f2c41ab… merkle_root: b83a7e21… tamper path, same exporter: provenance_status: integrity_failed ledger_broken_at: seq 22
M04 / COMPARE
What Continuity is — and what it costs next to the alternatives.
Memory platforms remember facts about users. Resume skills convert transcripts. Continuity records the work itself and can prove it. Different question, different layer.
| Capability | Continuity | Memory platforms (Mem0, Zep) |
Resume skills (free OSS) |
|---|---|---|---|
| Persistent memory across sessions | ledger-backed | vector / graph | summary files |
| Cross-model resume (Claude ↔ Codex) | proven, nonce-isolated | not the focus | transcript convert |
| Remembers the work itself — tasks, decisions, reversals | facts about users | what the summary kept | |
| Tamper-evident, hash-chained history | DB-enforced | ||
| Verified handoff receipt — computed, not asserted | verified / integrity_failed | ||
| Conflict detection with human resolution | first-class events | merge policies | last write wins |
| Local-first — no cloud required | one SQLite file | OSS core, paid platform | |
| Cost to start | $0 · open source | $19–$249 / mo tiers | $0 |
| Cost for a team | $24.50 / seat founding* | $125–$399 / mo+ | no team story |
*Target launch pricing — see M06. Alternatives per public pricing pages, 2026-07.
M05 / INSTALL
Install locally. Keep the ledger yours.
One package, one SQLite file, one MCP registration. Local SQLite, no cloud, no provider API keys for MCP/local workflows. Your ledger never leaves your machine unless you export it.
Install from PyPI
pip install continuity-mcp
Register with Claude Code
claude mcp add continuity --env CONTINUITY_DB="$HOME/continuity.db" -- continuity-mcp
Export a receipt
continuity-handoff --db "$HOME/continuity.db" --list-tasks
200+passing tests, including adversarial ledger attacks
3isolated persistence proofs: Codex→Codex, Claude→Claude, Codex→Claude
0provider API keys for MCP/local workflows
1file to back up: continuity.db
M06 / CLOUD
Continuity Cloud. For teams that bill for agent work.
CTY-CLOUD
MANAGED LEDGER
WAITLIST OPEN
One ledger. Whole team. Receipts for clients.
Continuity Cloud
The local ledger stays free and fully verifiable — Cloud is where your team shares it.
U1 · AGENCY → CLIENT
Deliver AI work with a receipt
Your team ships work partly done by agents. Attach a live, verifiable handoff link to the deliverable — the client sees what was done, who recorded it, and that nothing was altered. Reconstructing a dispute takes minutes, not a meeting.
WORK → LEDGER → RECEIPT LINK → INVOICE
U2 · PARALLEL AGENTS
Two agents, one repo, zero silent overwrites
Claude and Codex work the same project in parallel. When they disagree, the conflict is surfaced as an event with both positions on file — a human resolves it once, and the resolution rides into every future context.
CLAUDE + CODEX → CONFLICT SURFACED → HUMAN RESOLVES
U3 · MODEL ROTATION
Credits die. The task doesn't.
Hit a limit mid-task, switch to another model, or hand overnight work to a local model — the shared ledger keeps one thread across all of them, so nobody rebriefs and nothing regresses to a stale decision.
CLAUDE → LIMIT HIT → CODEX RESUMES → SAME THREAD
U4 · POST-MORTEM
Reconstruct any decision
"Why did we ship it that way?" Replay the chain: which output, which validation, which human approval, at which sequence. The answer comes with hashes, not recollections.
QUESTION → REPLAY → SOURCE ROWS → ANSWER
C1
Shared ledgersEvery agent, every teammate, one tamper-evident record. Concurrent sessions stop overwriting each other.
C2
Team attributionProcess-level attribution across teammates, agents, and sessions. Authenticated identity comes later.
C3
Client receipt linksSend a live, verifiable handoff link with an invoice instead of a PDF nobody can check.
C4
Priority supportDirect line to the builders, and founding members shape what ships first.
T1 · OPEN SOURCE
pip install continuity-mcp
$0forever
The trust layer is never paywalled.
Full local ledger — unlimited events
All MCP tools, cross-model resume
Verified HANDOFF.md export
Integrity verification, conflict events
Apache-2.0. Your machine, your ledger.
T2 · PRO — SOLO
Join the waitlist
$9/ mo target
Target pricing · waitlist
Everything in open source
Cloud hash anchoring — your Merkle root independently notarized
Encrypted ledger backup
Personal read-only receipt links
Your ledger stays local. We store hashes and an encrypted backup — never your plaintext work.
T3 · TEAMFOUNDING 100
Join the Founding 100
$49$24.50/ seat / mo
Target launch pricing. Founding rate locked for life.
Everything in Pro, shared
Shared ledgers with team attribution
Client-ready receipt links
Priority support + direct roadmap input
Founding members lock 50% off the Cloud launch price, for life.
Waitlist · Founding 100
Cloud should earn its build — founding demand decides what ships first, Pro or Team. Tell us which you'd pay for.
Founding members lock 50% off the Cloud launch price, for life. Pro and Team are target pricing until launch; the local product stays free either way.