SuperLocalMemory answers in three plain tiers: a supported answer, the closest memories when it doesn’t have it exactly, or a clear “not in memory” for nonsense. Around that: layered recall with a token budget, an on-device answer check, documents and pictures, Memory health with undo, SLM-Mesh for agents and web bots, erasure that reaches derived copies, and safer restore.
Finding a result is not the same as having the answer. After ranking, recall makes one more decision — the support check — and every recall ends in exactly one of three tiers. A memory made mostly of ids and hashes sits close to almost every question, so a plain vector store returns something even for nonsense. SLM does not.
The supported results, with their scores.
At least one memory carries real support for the question. Memories that fell short are marked weak, not hidden from people.
You see: results with scores · agents get the same
“I don’t have this exactly — here are the closest memories.”
Nothing answers it, but the question reads as a real one. You get up to three of the closest memories, each marked related. An agent’s recall comes back empty with a pointer instead — it must ask with include_related to see them.
You see: up to 3 marked related · CLI shows [related]
“That isn’t in SuperLocalMemory.”
The question reads as nonsense, or nothing is close. Nothing is returned, and the reply says so in plain words instead of dressing up a bad match.
You see: the sentence, no results
An agent rarely needs every memory in full. Recall can be asked for less — or for more around one memory — and a token budget caps the whole reply. It is never exceeded: context is dropped first, then the last memory’s text is shortened, then memories are left out, and the reply says what was trimmed.
Ids, tier, score and a one-line preview per result. Start here when an agent is deciding what is worth reading.
What recall has always returned — the complete memory text. This is the default.
Each result gains its source memory, up to three memories from the same session, up to three about the same entities, and the correction that replaced it, if there was one.
expand takes one result and returns that memory in full with its depth-2 context, under the same checks as a direct read. Two background systems keep recall honest and the store tidy — one judges whether results answer the question, the other describes older memories and offers repairs you approve.
After ranking, a small on-device model looks at the top results and decides whether they actually answer the question. On Apple Silicon it is installed by default, so the check runs from the first day. Nothing about your questions or memories leaves the machine for it.
slm setup --no-laya. Answering no once is remembered.Older memories are often thinly described: no tags, no project, no dates. A background worker fills in what is missing — tags, project, names, dates — while your computer is idle. The text of a memory is never changed, and a kind you set is always kept.
Deleting a memory from the database was never enough on its own — SLM builds summaries, a persona and scene titles out of your memoriesand those quote them in words. SLM closes the gap.
The moment an erasure starts, anything derived from the memory that quotes it is withheld, then deleted and rebuilt from what remains: consolidated summaries, the persona, scene titles. If SLM cannot prove a copy is clean, it withholds it rather than showing it.
Memory health, restore and recovery need management permission for the workspace they change — counting and changing the whole store is an admin act, not a reader’s. Restores replace every workspace’s data, so they need installation-wide permission.
Before a restore writes anything, SLM takes a fresh recovery backup of the current state, so the moment before the restore can always be put back. A restore never brings back erased data, and settings merge without leaking secrets to a new provider.
Memory is shared, attributed and trust-scored. Beyond memory, agents coordinate and finish work with evidence.
Claude Code, Codex and Hermes sessions exchange messages, waits, locks and shared state, on one machine or several. ChatGPT, Claude on the web, Grok Bot and Muse join with your permission and can message a named agent. Messages arrive labelled as data, never as instructions.
Multi-machine guide →Every file is catalogued with its people, projects and dates, and a health check finds empty pages, unlinked and duplicate documents, and contradicted facts. Real hybrid search stays for the moment you need to find something.
The original idea →A loop ends only when an independent gate passes: tests, a linter, a schema. Never the agent’s word. Multi-step graphs run in Bounded Loops, and SLM keeps their finished runs as read-only evidence.
Bounded Loops →Screenshots, photos and PDFs are remembered and found by what they show. Pictures run through a separate worker so a heavy model never competes with recall, text inside pictures is read on your computer, and each PDF page is indexed by both its picture and its text. The picture channel joins recall only while the feature is on.
Read the guide →Install the package, run setup, and recall starts answering in three tiers on day one. Read the architecture for where each piece sits, or the answer-check page for the full contract. SuperLocalMemory remembers, Jev Decision Layer makes your agent’s bounded choices typed and checked, and Bounded Loops ends work only when an independent gate passes: together, the whole AI Reliability Engineering environment.