Features
Five things that exist together nowhere else.
Individually, several of these have partial answers elsewhere. The combination is what does not exist, and the combination is what a team actually needs.
The handoff is a first-class object
A receivable artifact produced when work stops and consumed when it resumes. Not a record you have to know how to search for. A thing that arrives.
Everyone stores memory. Nobody hands off. The closest available thing is "query the memory store", which puts the entire burden on whoever is picking the work up.
Conflict resolution across humans and agents
Explicit arbitration when a teammate and an agent disagree about project state. Agent versus human-confirmed: the human wins, always, and the agent assertion is stored as disputed rather than silently applied. Human versus human is never auto-resolved.
Single-user products have no conflicts by construction. Products that do detect contradictions tend to invalidate the older fact automatically, which is exactly wrong for a decision, where a human has to arbitrate.
Provenance with actor attribution
Every item records whether a human or an agent asserted it, which one, when, and on what basis. That distinction is rendered everywhere it appears, because it is the distinction that decides what to trust.
Some products carry episode-level provenance for facts, or commit history. None distinguish human authority from agent assertion.
Selective rehydration under a token budget
Choose the minimal correct slice for the next task, with per-kind quotas so a pile of similar facts cannot crowd out every constraint. Load-bearing active constraints are always included, whatever the budget pressure.
Compaction and context editing shrink the window. They do not select for the task at hand. Compaction is lossy by design and task-blind, and semantic search returns what is similar rather than what is load-bearing.
Boundary-triggered structured checkpoints
Explicit capture at a task or day boundary into a typed schema, with contradiction detection before anything is written, and human confirmation on the load-bearing items.
Ambient capture produces noise. Threshold compaction fires when the window is full, which is the worst possible moment, and it produces nothing you can review.
Why it compounds
The value is not the feature list. It is the record.
After a year of checkpointing, your project carries its own history: what was decided, why, who confirmed it, and what was already ruled out. That record is yours, it is specific to your work, and it gets more useful every month it grows.
Questions
What people ask before they install it.
- What is Mneia?
- Mneia is the shared project memory and handoff layer for teams working with AI agents. It does three things: checkpoint, which captures the decisions, constraints, and open questions out of a session at a task or day boundary; rehydrate, which assembles the minimal high-signal context slice for the next task under a token budget; and handoff, which produces a receivable artifact when work changes hands.
- How is Mneia different from an AI memory tool?
- Memory tools give you a place to store context and a way to query it, which leaves the burden on whoever is picking the work up. Mneia produces a handoff: an object a person or an agent receives, stating what is done, current state, open questions, constraints, and the single next action. It also carries a superseded section recording what was tried and rejected, so a fresh agent does not propose the approach the team already ruled out.
- Why is context compaction a problem?
- Compaction is lossy by design and task-blind. It fires when the context window is full, which is the worst possible moment, and it produces nothing you can review. An agent that has been compacted can silently lose a constraint established two hours earlier and confidently propose an approach that was already rejected.
- What happens when a human and an agent disagree?
- The human wins, always. An agent assertion that contradicts a human-confirmed item is stored as disputed rather than silently applied. Two humans disagreeing is never auto-resolved, because arbitrating between two people is not a decision software should make on their behalf.
- Which AI coding tools does Mneia work with?
- Mneia ships an MCP server that works in Claude Code, Cursor, Codex, or any MCP client, plus a CLI and file interop with the AGENTS.md and CLAUDE.md files you already keep. The inner loop stays in your terminal, because a handoff that only works inside one vendor tool is a session feature rather than a handoff.
- Does Mneia replace my agent framework or observability stack?
- No. Mneia sits beside the frameworks and observability tools you already run, never above them. It is not an agent orchestrator, not a tracing or evals product, not an enterprise document search, and not an agent of its own. Everything it builds serves checkpoint, rehydrate, and handoff.