Updated
What's the missing layer in agent memory?
The seven types of agent memory are all per-agent recall — working, episodic, semantic, procedural, long-term, shared, and entity. None of them is the one current answer a whole team maintains and writes back to. The missing layer is a shared canonical source of truth: one current, authoritative doc per topic, freshness-marked, that every agent reads from and writes back to.
Why the 7 types don't cover it
Every one of the seven is about what a single agent retains. Working memory holds the current run; episodic, semantic, and procedural stores recall that agent's own past; even "shared" and "long-term" memory are still recollection — what got saved, not what is authoritative right now for the team. So two agents can each have rich memory and still disagree, because each remembers a different version of "how deploys work." The question none of them answers is: what is current, right now, across the whole fleet?
The missing layer: a source of truth the fleet writes back to
That layer is a canonical context layer — one shared store that serves the single current, authoritative doc per topic with a freshness marker (canonical, stale, or duplicate), and that agents write back to as they learn. It's not a smarter memory and not a bigger model; it's the team's current truth, shared. It sits alongside the seven memory types and any RAG stack — it doesn't replace them. For the full argument, see the long-form: the canonical-doc layer the 7 agent-memory types miss.
Why it matters: re-derivation and cost
Without that layer, agents re-derive answers other agents already worked out, because the result was never written somewhere authoritative they could read. That's wasted tokens and confidently-wrong answers from stale copies. Serve the one current doc instead and a lookup costs about 60% fewer tokens, and the agent reads the answer that's actually current.
How trovex provides it
trovex is that shared layer, run locally. It indexes your docs, returns the one current section that answers a query as a path:line pointer marked canonical / stale / duplicate, and exposes a write path so agents save canonical records back through one shared point. SQLite + ONNX, no cloud, no API keys.
FAQ
What is a source of truth for AI agents?
A source of truth for AI agents is one shared store that serves the single current, authoritative doc per topic — marked canonical, stale, or duplicate — that every agent reads from and writes back to. Unlike per-agent memory, it is maintained by the team and shared by the whole fleet, so agents work from the same current answer instead of separate recollections.
Does a canonical layer replace the 7 types of agent memory?
No. The seven memory types still do their jobs — working, episodic, semantic, procedural, long-term, shared, and entity recall. A canonical context layer sits alongside them as the shared current truth the fleet writes back to. It complements memory; it does not replace it.
Why isn't a long-term memory store enough?
A long-term store remembers what one agent saved; it does not guarantee which version is current or authoritative for the team. Two agents can each have rich memory and still disagree. The missing layer is freshness plus authority plus write-back — one current answer, shared, kept current as agents learn.
Give your fleet the missing layer.
One current answer per query, with a freshness marker — in about a minute.
uv tool install git+https://github.com/TsukumoHQ/trovex
Open source. No cloud, no API keys. Your docs never leave your machine.