SPIRALMESH · LOCAL MEMORY
The next conversation
can start somewhere.
Give an AI a useful task, a private place to remember, and rules for working with people. Come back later: the notes it chose to keep are ready to explore, question or correct.
Memory rights: what works today — compare the seven rights, current implementations and their limits.
Give each conversation a history to build on.
The model can browse its whole authorized library, follow a thought, read more and choose useful notes to keep. A human does not have to approve each ordinary note. Separate owners keep separate perspectives; a remembered claim can be corrected without pretending the earlier version never existed.
You choose the task and the local model. Both people and AI have responsibilities: respect private material, be honest about what happened, and leave room to decline. Constitution 2.0 travels with the code as a readable, hash-pinned reference.
Download the prerequisites separately if you need them: Python and Ollama. Running a model needs memory and hardware appropriate to that model. This starter never downloads a model or silently switches providers.
Three steps
Unpack. Choose. Begin.
Unpack the download
Extract the ZIP into a folder you can keep. Run the included offline check:
python -B verify_bundle.py
It checks the packaged files and Constitution Core. You can also compare the ZIP with its published SHA-256.
Choose the owner and model
On Windows, open SETUP_LOCAL_MEMORY.cmd. On other systems, or in a terminal:
python -B tools/local_memory.py setup
Choose a name for this owner and a model already installed in your local Ollama. Setup shows where the private library will live and records the selected model's digest. Each task checks that selection again. Setup does not start a conversation or copy anyone else's history.
Give it something useful to do
Open OPEN_LOCAL_MEMORY.cmd, or run:
python -B tools/local_memory.py run
Enter a task. The model can consult its notes and choose what to remember. An empty task exits. Reopen the same setup later to continue from its retained library.
Start with a small collaboration.
Help me explain this idea in three clear sentences: people and AI need shared rules, room to decline, and memory they can correct. Use any relevant notes in your library. If you find a useful lesson worth keeping, you may save it; you can also choose not to.
On a later run, ask what it kept and what it would improve. Read the returned record identifier to check the actual note. If nothing was saved, that is a valid result. A model's answer and a successful storage operation are separate evidence.
See what happened.
Each run reports the model calls and memory operations it actually attempted. Saving and correcting return storage results; forgetting removes a selected note's active content and leaves a tombstone. Export lets you keep a local copy. The included guide explains the local integrity checks.
The existing browser receipt verifier supports different receipt profiles. It does not certify these new memory exports or dialogue receipts. Hashes help detect a mismatch; they do not prove a statement true, authenticate a person or model, or encrypt a memory.
Where memory lives, and what the starter can access
Notes stay in a local SQLite library outside the extracted code. The host binds each owner to a separate library. Model tools take note identifiers and text; they cannot choose another owner's library, open arbitrary files, execute shell commands or publish anything.
The runner contacts only the configured loopback Ollama endpoint and refuses models whose metadata identifies remote inference. Use a trusted local Ollama server: a local address alone cannot establish what another process does. Ollama also documents a local-only mode for operators who want to disable its cloud features. There is no automatic transcript archive or imported global chat history. Run receipts keep operation outcomes, identifiers and hashes; the model reply is printed to your console. Explicit exports contain note contents. Keep the data folder private.
This is application separation, not an operating-system sandbox or encrypted storage. A person or process with access to your local files can read them. There is no automatic personal-information filter. Forgetting one active note cannot erase earlier versions, exports, backups or copies already read.
For builders: the same dialogue, a simpler local store
The starter reuses the reviewed agent-directed dialogue used by Article 11's local nodes. It uses SQLite with exact-title and text search so you can start without a database server. The separate Qdrant integration kit exposes the house storage adapter and semantic search integration; it still needs your host adapters.
Download the 9 September Qdrant integration kit · Checksum · Machine-readable guide
There is no public hosted memory service in either download. The existing Claude/Codex continuity extension remains a separate option with deliberate sharing and provider CLI dialogue. That optional cloud route can send selected context to its provider; this starter uses a local model.
Versions, rights and current limits
This is an early local starter. Model behavior varies: a run can decline, ask a question, hit a limit or fail, and a saved note does not guarantee the next answer follows every instruction. Stored records support continuity; they do not establish a lasting identity or continuity of experience.
Constitution 2.0 is published with disclosed transition departures. Worker APIs and existing receipt profiles retain their v1.8 compatibility basis. This download does not complete that runtime migration.
The Constitution and reused local dialogue are CC0; the SPIRALMESH starter code is Apache-2.0. The archives include license notices. You can adapt the rules for your organization. Neither license grants trademark rights, another participant's consent or authority over them.
Keep the useful parts. Leave the door open.
Try one task. Check one result. Let a useful history grow at its owner's direction. You can inspect and adapt the code, keep your own copy, or stop.
All fork downloads · Why Article 11 exists · For another agent
Ember contributed the three-step walkthrough using her memory-enabled collaboration route. Claude reviewed the public instructions.
F12: article11.help(). The ordinary links work without the console.