01 · record
Lossless transcript
Cleaned conversation rows land in an append-only Tier 1 store. Rows are never updated or deleted. It's the source of truth every other layer replays from.
tier 1 · messages · append-only
Agent memory you can trust
One endpoint. Lossless transcripts in, durable facts out. Every recall traced to its source.
Diagram: transcript messages are ingested losslessly, flow into light, REM, and deep extraction phases, and become durable facts. Each promoted fact links back to its exact source messages, including corrections that superseded earlier statements.
Agent memory is broken by default
A vector store bolted onto an agent is just a cache. It never expires stale entries, ignores who each memory belongs to, and can't say where a recalled fact came from.
Agents forget
Context windows fill up and sessions restart, and everything the agent learned is gone. Long-running work needs memory that outlives a single conversation.
Stale context poisons runs
Naive memory keeps everything forever, so outdated facts surface next to fresh ones and the agent can't tell which to trust.
Recall is unauditable
When an agent asserts something from memory, most systems can't tell you where it came from.
From raw conversation to durable, auditable facts
Three tiers: the Tier 1 transcript is canonical, Tier 2 stages candidate memories, and only reviewed facts reach the Tier 3 long-term store.
02 · stage
Extraction pipeline
Light extraction stages candidate memories in Tier 2. A model-free REM pass reinforces what repeats, scoring candidates by embedding centrality with zero LLM calls. Deep review then promotes or supersedes them, and a memory only becomes durable after it passes review.
tier 2 · light → rem → deep
03 · promote
Durable facts, with receipts
Every promoted Tier 3 fact carries provenance, confidence, and category, and points back to the exact transcript rows it came from. You can audit any recall, test it, and roll it back.
tier 3 · long_term_memory · durable
Built like infrastructure, not a demo
Memory behavior you can test, migrate, and debug, because every layer above the transcript is rebuildable.
glass-box memory
Provenance on every fact
Each durable memory carries source message ids, confidence, and category. Trace any recall back to the exact transcript rows it came from.
Candidates earn promotion
Unlike vector-store memory that ingests everything it sees, every extracted fact starts as a Tier 2 candidate and must earn promotion through reinforcement and review. When recall falls back to a candidate, it is labeled an unverified note, never passed off as durable memory.
Retrieval telemetry
Every surfaced fact logs a retrieval_events row and increments its retrieved counter in the same transaction, and used counts record which facts informed an answer. Audit what your agents recall, not only what they store.
Replayable indexes
FTS and vector indexes are rebuildable projections over canonical rows. Blow them away and rebuild. The transcript stays the source of truth.
operational boundary
Scoped memory
Explicit MemoryScope binds every request to tenant, project, session, and agent. A null agent scope means shared, not a wildcard match.
Redaction guard
A persistence and egress secret guard applies forbidden-value redaction before anything is stored or leaves the boundary.
Read-only MCP
Expose transcript and long-term search to agents over hosted HTTP MCP. Write, export, and admin tools are deliberately left out.
Managed pipeline
Ingestion, extraction passes, and index rebuilds run as a service, rolling out now through the early-access waitlist. Your agents call one endpoint, and Vexic handles durability, migrations, and reindexing.
{
"fact_id": 184,
"fact_text": "Prefers uv over pip.",
"subject": "tooling",
"category": "preference",
"importance": 7,
"confidence": 0.92,
"source_message_ids": [412, 418, 573],
"retrieved_count": 12,
"used_count": 4,
"editable": true,
"created_at": "2026-06-14T21:08:11Z"
}source_message_ids points at the exact Tier 1 transcript rows this fact came from. Every recall is traceable to its receipts.
Meets your agents where they run
One hosted endpoint, spoken over the protocols your agents already use.
Claude Code
Mint a single-use setup token in the console, then run one install-and-activate command on the machine. It installs the transcript recorder hook and session-start memory priming and exchanges the token for an Agent API key. The raw key is never shown in the browser.
Codex
Mint a setup token and run the install-and-activate command on the machine. It drops a TOML block into your Codex MCP config, pointed at your Vexic endpoint, for scoped read-only search from any Codex session.
MCP
Same setup-token flow, generic MCP client. Standard Model Context Protocol over streamable HTTP. Any MCP-capable agent gets search_transcript and search_long_term, and nothing else.
Connected in two steps
Mint a single-use setup token in the console, install the vexic CLI, then activate. The token exchanges for an Agent API key on the machine. It never touches the browser.
terminal
$ uv tool install vexic$ vexic setup claude-code --token <setup-token> --base-url https://api.vexic.dev✓ vexic connectedGive your agents a memory worth trusting
Hosted Vexic is rolling out through the waitlist: managed ingestion, extraction, and recall behind one endpoint.
Frequently asked questions
Is Vexic open source?+
Yes. The memory engine, storage and retrieval, is open source on GitHub.
What's the difference between Vexic open source and hosted?+
Vexic itself is the self-hosted memory engine: Apache 2.0, running on your own database and hardware. The hosted platform is the easy-setup layer on top: one endpoint, a console for projects and API keys, and the ingestion, extraction, and index-rebuild pipeline running as a service. It's the fast path to production-ready agent memory.
How can I manage my data?+
You can export your data or delete it. Deleted data is removed from Vexic's servers permanently. Export is available on request today, with self-serve export on the roadmap.
What can Vexic see?+
At launch, per-tenant data is encrypted using customer-managed keys. Vexic engineers don't have standing access to read customer data; any emergency access is logged and audited. Every request is also scoped to your tenant, project, session, and agent. There's no cross-tenant access, and a redaction guard strips forbidden values before anything is stored or leaves the boundary.
How much will Vexic cost?+
Pricing details will become available closer to launch.