interactive-tui
Streaming terminal interface
Responses arrive live inside a focused TUI, so planning, tool calls, and code edits remain visible without leaving the shell.
Harness Agent / terminal-native LLM automation
MNEMS
A coding agent living inside your terminal: interactive TUI, persistent conversation memory, headless automation, remote sandboxes, custom skills, and human approval for every dangerous move.
$ mnems resume thread:deploy-gateway
streaming response... tools armed: shell, editor, sandbox, channels
streaming response... tools armed: shell, editor, sandbox, channels
Core feature matrix
Not a chatbot. A harness.
MNEMS wraps model reasoning around the real workflow surface developers already trust: the terminal. It streams, edits, remembers, resumes, and can run without a human screen when CI needs an agent.
resume
Persistent conversation memory
Thread state survives sessions, letting developers reopen work exactly where the agent left off.
-n / headless
Automation and CI mode
Non-interactive runs make MNEMS useful in scripts, pipelines, release checks, and scheduled maintenance.
skills
Custom extensions
Teams can define reusable skills for repo conventions, deployment recipes, code review rituals, and domain tools.
approval gate
Human in the loop
Tool calls can be reviewed, approved, or rejected before execution, keeping power and control in the same frame.
remote execution
Sandboxed code runs
Connect isolated environments such as LangSmith, AgentCore, Daytona, Modal, or Runloop for safer execution.
Live transmission layer
Motion makes the harness feel alive.
The videos become monitored signal feeds: agent wake state, memory recursion, and daemon pulse. They add cinematic energy without breaking the terminal-native product story.
stream / muted
loop / local
status / active
Terminal choreography
Every action leaves a trace.
MNEMS treats the command line as the cockpit: readable state, explicit approvals, inspectable logs, and direct handoff between human intent and LLM tool calling.
$ mnems "ship the gateway daemon health probe"
↳ loaded repo memory: channels, adapters, sqlite logs
↳ plan: inspect status command, patch probe, run tests
tool call pending: edit packages/channels/health.ts
approve? [y]es / [n]o / [d]iff
✓ approved by operator
✓ tests passed: channels-status, logs-tail, retry-state
$ mnems channels status --probe
telegram-adapter connected 12ms
gateway-daemon healthy sqlite:local
Multiple channels / gateway daemon
One agent, many surfaces.
A background daemon coordinates channel adapters, probes health, persists status locally, and gives developers a fast operational view from the CLI.
Telegram adapterconnected · retry policy armed
Terminal TUIinteractive session · streaming
CI runnerheadless task · non-interactive
Gateway
Daemon
Daemon
SQLite stateadapter status · local logs
Probe commandmnems channels status --probe
Logs commandmnems channels logs
Sandboxed execution
Let the agent move. Keep the blast radius small.
MNEMS can delegate risky code execution into remote sandboxes while preserving human approval gates for local tool calls.
01
Request
The developer asks MNEMS to investigate, patch, test, or automate inside the current repo.
02
Route
The agent chooses local terminal work, a remote sandbox, or a channel adapter based on the task boundary.
03
Approve
Destructive or privileged tool calls pause for review, giving the operator a clear accept or reject moment.
04
Resume
Conversation state is saved, so the same thread can continue after a terminal restart or a later shift.
Persistent memory is the sigil: thread state, local logs, channel status, and custom skills held together as operational context.
Memory that survives the window
Conversation resume keeps the agent from becoming a disposable prompt box. It can pick up an interrupted investigation with the same thread context.
Skills that encode team ritual
Custom skills turn repeated work into reusable operating procedures: review style, deploy sequence, log triage, migration rules, and more.
Local reporting that stays fast
SQLite persistence means adapter health and logs are available immediately from the CLI, even when channels are retrying.
Tool calling with a human brake
MNEMS can act, but sensitive operations remain inspectable. Approve, reject, or ask for a diff before the agent proceeds.
Install the harness
Bring model reasoning to the place where real engineering work already happens.
$ mnems -n "run release checks, summarize failures, and wait for approval before edits"