ptkachykby
Joined 11 days ago
Karma
2
Submissions
22
Member since
Oct 21
Recent submissions
Comment · 4 days ago
Comment · 4 days ago
One more thing that could help: memory is only context for Claude Code, so use a PreToolUse hook that blocks the command itself.
A hard part is that the hook blocks only 1 command and does not carry the reason to the next session.
On Substrate there is also 1 memory outside of any agents, where every fact has a source.
Did the saved note appear in the next session? https://trysubstrate.co/l/a3hn
Comment · 5 days ago
Also good to know about /doctor prompt-audit in Claude Code 2.1.283+:
It reports if there are any contradictions or outdated information in instruction files and does not change anything.
It seems to be hard because it compares the files with each other, not with decisions that were made in sessions.
In Substrate you can see what is in memory and select what each agent sees.
Does decoherence show first in instructions or in generated code?
Comment · 5 days ago
Also, a tip for Claude Code: put rules in .claude/rules and add `paths:` frontmatter, so a rule loads only for matching files.
A hard part of this approach is scoped rules: you still need a human to decide which records are still true.
What we try to do at Substrate with our memory is to automate this: 1 memory outside of any agent, where all the facts have their source.
Did the agent-written records go stale or just pile up?
Comment · 5 days ago
Each agent sees only the memory you allow it, per agent.
Did you have a chance to try it?
https://trysubstrate.co/l/a2hn
Comment · 5 days ago
We don't move session / skill files between tools.
Instead the next tool gets the decisions from the last session, as well as the sources they were based on.
The plugins are open source (MIT), the server isn't, and there's no self-hosted version yet.
Is it enough to have the decisions and open questions, or do you need the full transcript?
Comment · 5 days ago
Also, if you use Claude Code, it reads a gitignored CLAUDE.local.md next to CLAUDE.md, so you don't need a second repo. For worktrees, import one file from your home directory: @~/.claude/my-project-instructions.md
The hardest part is keeping the information true in all of these prompts as the project evolves. We built Substrate for that: one memory outside of any agent that Claude Code, Codex and Cursor read, where every fact keeps its source.
Do your agent files drift between tools? https://trysubstrate.co/l/a3hn
Comment · 6 days ago
You could also increase compaction.reserveTokens in ~/.pi/agent/settings.json to make pi compact more often.
Also you can use /compact <instructions> to specify what facts should be included in the summary.
One hard part is that from a handoff file you only have information for 1 session, so you have to manually determine next week which facts are still true. At Substrate we solve this problem using our memory: we have 1 memory outside of an agent and track the source of every fact. https://trysubstrate.co/l/a2hn
Comment · 9 days ago
The problem then is how to avoid a situation where 1 wrong fact is present in the common memory and then propagated to all agents. So you need to be able to see what's in the memory, correct it once and then make sure this change is propagated to all agents.
This is 1 of the ideas we used when we were building up memory in Substrate.
How do you store the shared memory?
Comment · 9 days ago
A markdown file works well as long as you don't use a second agent / tool that needs access to this same file. Otherwise, you will find that the copies of the file drift apart.
It's better to store the memory outside of all tools and then connect all of them to it via MCP. Then you have a single store where all of the tools can read from and write to.
This is one of the ideas we used when building up memory in Substrate. How many tools are using this file as memory?
Comment · 9 days ago
You can add a reference to the source session for each note in the vault.
Then you could check if this is true in the actual session instead of trusting what the note says.
This is one of the ideas we used when we were building memory for Substrate.
Does your agent write to the vault automatically?
Comment · 9 days ago
If the skill wa already loaded into context tranplanting it into a new harness will move that skills context as well (not the scripts though).
What kind of memory has degraded your performance? Memory is quite tricky to tune and simple a direct implementations of an LLM Wiki concept can indeed have a lot of problems, as context accumulates.
Substrate, for example, keeps the history of edits to every fact, and the source of it.