roadmap
Where Mnemo is going
What is being built, what comes next and what will not happen. It is a statement of intent, not a promise: there are no dates on purpose, and the order changes when real use says it should.
Coming next: open for review
Written and tested, in a draft pull request that has not been merged. Not in any release, and it may change in review.
Hooks
Run your own commands before and after tool calls and sessions. A hook can block a call, and no hook can loosen the permission gate.
Schedules
mnemo schedule runs a prompt on a timer. Jobs live only in your own ~/.mnemo/schedules.json, never in a project, and a run is read-only plan mode unless you gave the job actions.
Language-server diagnostics
After an edit, the agent is told about the errors your language server reports. Opt-in through lsp.json, and only in a folder you trust.
.agents/ as the project folder
A project's own settings, skills and MCP config move from .pi/ to .agents/. Breaking for a repository that only has .pi/.
In progress
Being built or tested now.
Background curator
While you are away, it uses your default model to tidy memory, merge duplicate facts, resurface forgotten knowledge and propose skill changes. Facts apply on their own; skills wait for your review. It runs only when idle and within a daily token budget.
Remote view
Watch and answer a running session from a browser. Loopback only, behind a token, and off by default. Being built; no pull request yet.
A marketplace for skills and MCP servers
Read-only, and every entry pinned by content hash so what you install is what was reviewed. Being built; no pull request yet.
A pass in real terminals
Every screen checked by eye in iTerm2, Ghostty, Windows Terminal and a 16-colour fallback, with screenshots in the docs.
Next
The work most likely to land in the next releases.
A text index behind grep
The code index finds definitions and calls today; the plan adds a trigram text index so grep is fast on a large repository too.
Memories pinned to definitions
A fact can point at a function and go stale when that function's body changes.
Call-graph ranking
Order code search results by how much of the project calls them, not only by words.
The VS Code extension on the Marketplace
It works from a downloaded .vsix today; publishing it means it installs and updates like any extension.
The memory claim, measured again
The first result (memory 3/3, no memory 0/3) is three tasks on a free model. The aim is a larger set of repeated sessions where the second run is the measurement, published with its method.
A review of the curator against real models
It has been tested with scripted models only. Before it ships it needs weeks of real use and a revert path you have seen work.
More agents' transcripts for ingest
mnemo memory ingest reads Claude Code sessions today. Codex and opencode transcripts would let the memory learn from them after the fact.
Skills that cite their evidence
A skill change should point at the sessions that justified it, and a skill that stops helping should be retired.
A clean-machine install test on every OS
A person follows the README on a fresh virtual machine per system, with nothing preinstalled.
Later
Wanted, not yet scheduled.
Trace tools
A way to read the redacted per-call logs: mnemo traces.
A sandbox
Decided: an OS sandbox is deliberately skipped for now, because it needs a native helper per system that cannot be tested here. The permission gate filters; it is not a sandbox, and the docs keep saying so.
A retrieval upgrade
Switch the memory sidecar to its approximate-nearest-neighbour path when a project's memory grows past a size where it pays.
Merged, in the next release
On the main branch and tested, but not yet in a downloadable release (the newest is 0.1.8).
A code index for the agentmain
code_search, code_symbol, code_refs and code_map answer where something is defined and what calls it, and after an edit the agent is told what elsewhere calls what it changed. Two settings, codeIndex and codeImpact, turn it down.
Commands get an allowlisted environmentmain
bash and the Python kernel no longer inherit your whole environment, only a short list. passEnv names more. MCP servers over stdio get the same list, and MCP over HTTP stays off unless your own mcp.json says allowNetwork.
Managed policy and network egress rulesmain
A policy file user and project settings cannot widen, and outbound rules that refuse cloud metadata addresses by default.
Deny rules hold inside the Python toolmain
A path the gate refuses in bash is refused in a Python cell too.
Plan mode asks before it plansmain
Clarifying questions with choices, answered before a plan is written.
--model, and a provider's preferred modelmain
Pick the model on the command line. Bedrock is no longer the default because one AWS variable is set, and a provider starts on its preferred model.
Stop at the first 401, 402 or 403main
A refused key or an empty balance ends the run with the reason instead of retrying.
mnemo -p says where it worksmain
When the folder has unfinished work or no git, a headless run says so up front.
Shipped
Done, in the release named.
A VS Code extension, MCP servers as tools, a wider /rewind, a prompt budgetv0.1.8
mnemo --rpc for editors, MCP through the permission gate, shell changes rewound, a dollar cap per prompt, and nothing a folder chose runs before it is trusted.
Agents that talk to each other on their ownv0.1.6
A message wakes an idle agent, which can answer; bounded at 12 messages in a row and a $0.50 allowance for turns messages start, until you prompt again.
Type to a running sub-agent; links that survive a restartv0.1.6
A focused sub-agent pane in the split is a prompt to it, and linked projects are linked again when both are open.
The playground's conversation viewv0.1.6
One pair of agents' messages as a thread.
Agents in teamsv0.1.5
Sub-agents that start sub-agents, sibling agents that see and message each other, split layouts and the playground.
Permission holes from v0.1.3 closedv0.1.4
An approval no longer covers a subcommand that runs something else, and plan mode does not write memory.
Login, memory and Windows install fixes from a full reviewv0.1.3
Subscription login in the downloaded binary, memory that stays on, a Windows exe that installs itself, about sixty fixes.
Install, update and uninstall commands, animated installersv0.1.2
mnemo update, mnemo uninstall --deep and installers that draw themselves.
Provider keys kept out of agent commands; a bounded Python kernelv0.1.2
The model's commands no longer inherit your API keys.
Rewind, with a checkpoint before each editv0.1.2
Go back to an earlier message, the files, or both.
Trust-weighted recall, memory tidy, edit in the memory pane, your own embedderv0.1.2
What helped ranks higher; duplicates and stale facts get retired.
ctrl+k palette, ctrl+f search, /resume search, /stats, clickable links, notificationsv0.1.2
Find and reach things faster.
Eleven themes and full-terminal animationsv0.1.2
Mnemo Night to Paper, filling the whole window.
Engram in more placesv0.1.1
Boot, welcome card, empty transcript.
The first public releasev0.1.0
Memory, skills, kernel, sub-agents, best-of-n, escalation, several agents at once, four permission modes, one-line installers.
Not planned
Said plainly so nobody waits for it.
Telemetry
There is none and there will not be. The only network traffic is your model provider and what you start yourself.
Hosted memory
Memory stays on your machine. Syncing it through a service we run is not the plan.
The previous Go and Node stack
It lives on the legacy branch and is not maintained.
Want to shape it? Add a thumbs-up reaction to an issue, or open one. Every release's changes are in the changelog.