Money
Recall costs zero tokens. The same fact re-reads forever, and keepsakes restore a full agent with one command.
MNEMOS
Try it
Most agents lose everything the moment the tab closes. Mnemos keeps memory durable, gates money on it, and refuses to act on a memory that turned out wrong.
$ mnemos revise preference contractor_rate 60 --reason "renotiated"
revised preference:contractor_rate: 40 -> 60
suspect: agreement:fencing
$ mnemos pay fencing 160
gate: agreement 'fencing' is suspect: depends on revised memory
contractor_rate; reconsider first
payment refused
Every agent has the same correction story. You say the rate is wrong, the agent says noted, and the payment still goes out. Mnemos computes the blast radius of every correction, marks dependents suspect, and the gate refuses until each one is explicitly reconsidered.
revise the fact, history stays
dependents become suspect
the payment gate refuses
review reopens the gate
"a correction that does not propagate is a new fact sitting next to an old decision. walking every agreement built on the wrong number is the part nobody builds because it is tedious. tedious is where the money leaks."
Sibyl Cap, on the revision gate
Recall costs zero tokens. The same fact re-reads forever, and keepsakes restore a full agent with one command.
The gate refused a real payment while the agreement was only agreed, then executed it once delivered. Memory alone changed the outcome.
A high severity lesson vetoes linked payments until resolved. Mistakes stop repeating.
When a memory turns out wrong, everything built on it becomes suspect until reviewed. Nothing deleted, all journaled.
Background consolidation over the journal. Repeated patterns, unanswered questions, and refused payments become review-gated proposals. Nothing applies itself.
Memory at any past moment. Walk the revision history backwards and diff what changed: what did we believe, and when did it change.
Deterministic entity extraction on every fact. Recall boosts facts that share an entity with the question. Zero model calls.
FTS first, semantic assist when you bring a key. An optional embedder fuses cosine with keyword search, off by default.
A bounded proactive queue grounded in memory. The most urgent matter surfaces, one per tick, and declined matters stay declined.
A curated summary of who you are: identity, preferences, accepted principles, standing agreements. Not a raw history dump.
Seeded, reproducible runs of the real gate. 12 trials, seed 1337, no fabricated numbers.
Reproduce with scripts/ablation.py. Pinned by tests.
Full captures and reproduction for both partner stacks: docs/VERIFICATION.md.
One config line. Your agent gets durable memory, the payment gate, and revision.
{
"mcpServers": {
"mnemos": {
"url": "https://mnemos-production-2572.up.railway.app/mcp"
}
}
}
Runs on the hosted endpoint. No install, works on any machine.
{
"servers": {
"mnemos": {
"type": "http",
"url": "https://mnemos-production-2572.up.railway.app/mcp"
}
}
}
Streamable HTTP, works straight from the mcp.json file.
https://mnemos-production-2572.up.railway.app/mcp
Works with ChatGPT connectors, Claude web, and any streamable HTTP client. One click install on Smithery.
Real memory, real tools, and an agent model answering. Chips run raw tools, the box is a conversation.
Private memory: each browser gets its own store, and reopening keeps it.
Connect the MCP endpoint and Mnemos remembers, gates, and corrects for you.