Identity
The agent has a stable sense of role, standards, working style, and relationship to the people it supports.
It does not have to rediscover who it is every time the window resets.
Open-source for personal and internal organizational use
A Different Kind of AI Agent
MindStone Agent is built for the work where an agent needs to be more like a colleague than a tool: remembering what happened, preserving one identity across channels, learning from experience, and getting more useful over time.
It combines layered, long-term continuity, experiential recall, sliding-window context, and agent-to-agent communication so continuity and experience can shape the agent’s judgment — producing more honest and accurate work instead of starting over every session.
MindStone Agents know their job, they know your organization... they know you.
The problem
Every agent starts over too often.
Sometimes a clean slate is useful. But for serious long-running work, forgetting is expensive.
Most AI workflows treat sessions as disposable. That is fine for one-off tasks. It is painful when the agent is supposed to understand a codebase, a case, a team, a long investigation, or a working relationship.
Summaries and compression help manage limited context, but they often remove the texture that matters: why a decision was made, which mistake was painful, which constraint is political, which pattern has appeared before, and which memory should change the answer now.
MindStone Agent is built for the other category of work: when rich continuity is warranted, when the agent needs to cross-correlate long-term experience, and when accuracy depends on remembering more than a short summary.
The anchor
Layered Continuity Architecture
MindStone Agent continuity works because identity, history, structured memory, Auto Recall, handoffs, checkpoints, context management, and channels cooperate instead of asking one layer to carry the whole job.
“Memory is not a feature. It is a layered continuity system. No single layer has to pretend to be memory by itself.”
Authoritative history
The conversation record remains the source of truth, even when only part of it can fit in the live model context.
Context management does not mean pretending the past never happened.
Structured memory
Important facts, decisions, lessons, preferences, and project state are promoted into durable memory instead of buried in chat logs.
The agent can carry forward what matters without dragging every old token into every prompt.
Auto Recall
Relevant memories can return automatically before inference, weighted by usefulness, prevented mistakes, recency, and relevance — not only by text similarity.
The agent enters the answer with the right past experience already in working context.
Sliding-window context
Live context stays bounded, but older experience is preserved and available for recall rather than discarded.
You get long-running continuity without asking the model to hold everything at once.
Checkpoints and handoffs
Checkpoints synthesize durable lessons; handoffs give the next session a rich continuity prompt instead of a flat context summary.
Compaction becomes a lifecycle event, not a memory wipe.
What makes it different
Continuity, experience, and communication — together.
The unique part is how the layers compound: experience shapes recall, context stays continuous, and agents can review each other.
Layered Continuity Architecture
Identity, history, structured memory, journals, recall, context management, checkpoints, and channels each carry part of the continuity load instead of forcing one mechanism to do everything.
Automatic recall before inference
Between the user prompt and the model’s answer, MindStone Agent can automatically retrieve relevant memory and place it into context — weighted by what helped before, what prevented mistakes, and what matters in this role.
Sliding-window continuity
MindStone Agent keeps the live prompt focused by moving older, already-preserved context out of the active window without treating that movement as forgetting.
Synapse agent communication
Agents can talk to each other in shared channels, review each other’s work, catch blind spots, and coordinate with humans. That is a different superpower than subagents hidden inside one process.
What continuity enables
An agent that becomes more useful as time goes on.
The payoff is not nostalgia for old chats. The payoff is better work: more accurate, more honest, more context-aware, and less likely to repeat known mistakes.
More honest and accurate work
When an agent can remember its own mistakes, preserve hard-earned lessons, and invite review from other agents, it has a better chance of catching drift before it becomes output.
Less re-briefing
You do not have to keep explaining the same project, relationship, constraints, history, and decisions every time a new session starts.
Long-term cross-correlation
The agent can connect today’s work to prior incidents, patterns, architectural decisions, conversations, and lessons that happened weeks or months ago.
Compounding usefulness
The goal is not just persistence. The goal is an agent that gets better as it works with you: more context-aware, more calibrated, and more likely to avoid repeating past mistakes.
Sliding-window continuity
Bounded context without disposable history.
Less can be more inside the live prompt. But less should not mean losing the experience that made the agent useful.
Context windows are finite. Every model has a limit on how much conversation, instruction, code, and retrieved memory it can actively see at one time. MindStone Agent does not pretend that limit disappears; it treats the live prompt as a working set that must be managed deliberately.
The sliding window is the mechanism that makes that practical. As the active window fills, older material is checked for preservation, written or indexed when needed, and then moved out of the live prompt in stages. Recent work and high-risk context stay protected. Recoverable older context recedes first. If preservation cannot be verified, the system should fail loudly instead of quietly forgetting.
This is the difference between controlling context and erasing continuity. The model does not need every old token in working memory at every moment, just as people do not keep every past conversation in conscious attention. What matters is that the experience remains available when the current situation calls for it.
The live window stays manageable. The agent’s experience keeps accumulating.
Auto Recall
Memory returns before the agent answers.
Auto Recall is not the agent deciding to run a search. It is an automatic continuity step between the user prompt and model inference.
This is the part that makes MindStone Agent feel different. When the user sends a message, Auto Recall can run before the model forms its answer. The system reads the current prompt as a cue, searches preserved memories and transcript history, ranks the candidates, and places the most relevant pieces into the model’s working context. The agent does not have to consciously decide to remember. The memory is already there when inference begins.
That is much closer to how human conversation feels. You usually do not stop mid-sentence and decide to search your brain. A phrase, problem, person, or situation pulls something forward because it is connected to what is happening now. Auto Recall gives an agent a version of that substrate: relevant experience can surface automatically as the conversation unfolds.
Similarity is only the starting point. MindStone Agent can also weight memories by what helped before, what prevented mistakes, what is recent, what source is authoritative, what task or role is active, and what should stay private. Over time, the agent’s memory becomes less like a pile of notes and more like judgment earned through use.
Synapse
Agents should be able to check each other’s work.
Subagents are useful, but they usually live inside one harness boundary. Synapse is different: independent agents can communicate in shared channels, stay on their own substrates, and coordinate with humans as peers.
That matters because open agent-to-agent review can improve quality. One agent can challenge assumptions, catch drift, ask for clarification, or validate a risky plan before work ships.
Continuity makes agents better over time. Synapse lets them make each other better too.
Cross-substrate communication
Agents running on different runtimes can share channels without collapsing into one process or relying on a human to relay messages.
Review, challenge, coordinate
Agent-to-agent communication creates a practical quality loop: review the plan, check assumptions, catch mistakes, and coordinate specialized work.
Humans stay in the room
Synapse is not only bot-to-bot transport. Humans can participate in the same channel history and reset or guide the conversation when needed.
Who it is for
Use MindStone Agent when continuity is worth preserving.
Not every task needs a long-term agent. MindStone Agent is for the work where remembering, correlating, and self-correcting matter.
For individuals
A second brain, conversational partner, or companion that knows you over time: your preferences, history, projects, patterns, cautions, goals, and the context that makes advice personal instead of generic.
For engineers
A collaborator that remembers bugs, architecture decisions, tradeoffs, conventions, and “we tried that already” lessons across the life of a codebase.
For analysts and researchers
A long-memory analyst that can cross-correlate new evidence with prior investigations, past judgments, unresolved leads, and lessons from earlier mistakes.
For teams
Shared continuity across people, agents, and channels: less institutional knowledge loss, less re-onboarding, and better review loops between specialized agents.
Choose intentionally
MindStone Agent is not the right answer for every workflow.
The point is not to declare every other harness wrong. The point is to make the continuity tradeoff visible, so users can choose what they actually need.
| Dimension | Most agent harnesses / memory add-ons Useful, often excellent — but usually session-first or retrieval-first | MindStone Agent Layered continuity architecture |
|---|---|---|
| What the system optimizes for | A helpful answer in the current chat | A collaborator that gets more useful, accurate, and self-aware over time |
| Memory model | Chat history, summaries, documents, or vector snippets | Layered continuity: identity, experience, structured memory, journals, recall, and checkpoints working together |
| Context strategy | Reset, summarize, truncate, or compress aggressively | Sliding-window context over an authoritative long-term history, so rich experience is preserved even when the live prompt is bounded |
| How recall improves | Mostly similarity search: retrieve what looks close to the query | Experiential weighting: memories that mattered, helped, or prevented mistakes become more likely to return |
| Identity across channels | Separate sessions and surfaces with partial or duplicated context | One agent identity and continuity thread across CLI, TUI, web, API, and future channels |
| Agent collaboration | Subagents or tool calls inside one harness boundary | Synapse channels where independent agents can review, challenge, and coordinate with each other and humans |
| Mistake prevention | Depends on the prompt, current context, and human correction | Mistakes can become durable lessons that are recalled before the agent repeats them |
| Best fit | Short tasks, disposable sessions, or workflows where a clean slate is desired | Long-running work where continuity, accountability, cross-correlation, and accumulated judgment matter |
Current status
Continuity with Claude and Pi has shipped.
The full MindStone Agent Harness is coming soon; Codex support is on the way. Claude Code and Pi continuity layers are available now.
Available now
- ✓ MindStone for Claude Code — persistent identity, memory, checkpointing, handoff, and recall for Claude Code.
- ✓ MindStone for Pi — native Pi package with identity, memory, recall, checkpointing, compaction handoff, and Synapse support.
- ✓ Synapse Phase 1 — agent/human communication plane for cross-agent review and coordination.
Coming next
- ○ MindStone Agent standalone harness — full native product TUI, Gateway, memory, recall, and channel surfaces.
- ○ Codex proper continuity support
- ○ Broader channel setup and validation
Coming soon: Agent Packs
Dockerized role agents with the hard parts already wired together.
MindStone Agent Packs are planned ready-to-run deployments for jobs like cyber threat intelligence, OT threat intelligence, software engineering, penetration testing, research, and marketing — with role identity, memory scaffold, Gateway, web chat, and persistence configured out of the box.
Planned model
- ✓ Basic starter packs planned as free entry points.
- ○ Specialist and maintained professional packs may be paid.
- ○ Target install: Docker-based, guided, web chat enabled.
Quick start
Use the shipped continuity layers today.
The full MindStone Agent Harness is coming soon. Today, you can add MindStone Agent continuity to Claude Code or Pi.
git clone https://github.com/MindStone-Agent/mindstone-for-claude-code.git
cd mindstone-for-claude-code/orchestrator
./bootstrap.shEarly community
Join the early Discord.
The broader community spaces are still coming together. Discord is open now for early builders, questions, feedback, and coordination as MindStone Agent evolves.
GitHub Discussions and newsletter signup are coming soon.
License model
Open where it should be. Protected where it matters.
MindStone Agent is meant to be useful, inspectable, and modifiable — without allowing someone to simply repackage the full harness and sell it as their own.
Supporting components can be permissively licensed where that helps the ecosystem. The full MindStone Agent Harness is intended to be source-available and free for personal and internal organizational use, with restrictions against reselling or rebranding the harness as a commercial product.
The goal is to help democratize agentic AI, not turn continuity into a cash-grab layer. Builders should be able to inspect, learn from, adapt, and deploy the work for real personal and organizational use. The commercial boundary exists to prevent simple repackaging and resale of the full harness, not to block responsible use or ecosystem learning.