Shared Memories for Agent-Powered Communities
← Peter Le Bek
Shared Memories for Agent-Powered Communities
August 2026
Coding agents are surprisingly good at learning undocumented APIs.<br>The problem is they usually forget what they learned when the session<br>ends.
For one user it’s easily solved with an AGENTS.md or<br>folder of notes. For a team just commit those notes to Git.
But what if thousands of people are independently exploring the same<br>undocumented API with their agents?
I ran into this problem while building Modmixer, a coding agent for modding<br>games. Modding involves reverse-engineering binaries, finding injection<br>points, and learning undocumented runtime behavior through<br>experimentation.
With many people modding the same games, I wanted a discovery made by<br>one person’s agent to become useful to everyone. Instead of each user<br>spending tokens and compute rediscovering the same things, the community<br>should gradually build up a shared memory of everything its agents have<br>learned.
In practice, this means turning discoveries into small, reusable<br>findings that make sense for the domain. In Modmixer, each finding has a<br>title that says what was learned, a<br>recipe with the working code or steps, and a<br>“why it’s tricky” line describing the plausible wrong<br>turn an agent would otherwise take.
Once you have a useful unit of knowledge, the rest of the system is<br>simple:
CLIENT 1..N
AGENT
WRITE
LOCAL CACHE
submissions -->
UPLOAD
SUBMISSIONS
skill -->
REVIEW
AGENT REVIEW SKILL
curated -->
ACCEPT
CURATED
client -->
RETRIEVE
FIG. 1 — SHARED KNOWLEDGE SYSTEM
The server has two database tables: submissions and<br>curated.
When an agent makes a discovery, it gets a unique ID, is uploaded to<br>submissions, and is cached locally so the agent can use it<br>immediately.
Submissions are untrusted. An agent skill periodically reviews them<br>and inserts accepted findings into the curated table. It<br>can:
Verify claims with searches or tests
Detect malicious content (e.g. prompt injection)
Merge duplicates and resolve conflicting findings
Normalize formatting
Escalate ambiguous cases to a human
Each curated finding records the IDs of the submissions it was<br>derived from, allowing clients to remove superseded findings from their<br>local cache.
The result is a shared memory that gets better as more people use<br>it.
Follow me on X