AccInt

A model that scores 90% on a benchmark today scores 90% tomorrow. It doesn't learn from deployment, doesn't track which of its outputs led to good outcomes, and doesn't remember last week's mistake. It generates intelligence and throws it away. You keep paying — in time and tokens — to rediscover what already worked.

Accreted Intelligence is a bet that this is temporary. The idea is to move learning out of model weights and into scored external state, where judgment compounds from contact with reality and the model becomes a replaceable processor rather than the place intelligence lives.The reasoning engine is the part you can swap; the judgment it earned in your world is the part you keep.

There's a gap in how the existing tools are positioned, and it's whereaccsits. Memory remembers context. Observability shows traces. Automation runs playbooks.acccloses the learning loop: commitment, action, approval, outcome, reusable path — scored by results, audited on a ledger, and running fully on your hardware. And it does the one thing memory can't: itpredictsthe path most likely to work from everything that worked in your world before, then watches its own error.

accis a working kernel for that thesis. It's a Recursive Language Model over a late-interaction scored-token memory: two verbs over one memory. Credit defaults to a weak prior, and only reality earns full weight.

There is no separate mode for technical and non-technical work. The loop is identical whether you're shipping code or chasing invoices. Only the content of what's retrieved and acted on differs. You talk to your agent in plain words, and the domain lives in the content rather than the architecture.

Four different jobs, one set of primitives:commitment → action →HELD → your OK→ outcome → credited lesson.The authority gate (HELD → your OK) is structural in every flow that touches the outside world. That gate is what makes the same loop safe for consequential work and not only for code.

Run it again next week and verified steps replay instead of re-reasoning. Most AI re-reasons every task from scratch, so you pay full price forever.accpredicts the path that worked and replays the verified steps, so the same job costs less every run and keeps dropping as it learns.

accinstalls in one line. It runs the installer for your OS, which setsaccup on your machine:

curl -fsSL https://raw.githubusercontent.com/maxbaluev/accreted-intelligence/main/bootstrap/install | ACC_INSTALL_REF=github-readme ACC_INSTALL_SOURCE='ref=github-readme&utm_source=github&utm_campaign=readme' sh

$env:ACC_INSTALL_REF='github-readme'; $env:ACC_INSTALL_SOURCE='ref=github-readme&utm_source=github&utm_campaign=readme'; irm https://raw.githubusercontent.com/maxbaluev/accreted-intelligence/main/bootstrap/install.ps1 | iex

The installer probes your hardware, picks the embedder tier it can honestly run, downloads and verifies the matching release binary when available, starts a warm local daemon, and wires your agent's.mcp.json. The first run may download the embedder model (several GB) and take minutes. The installer reports the wait honestly and never pretends your hardware is bigger than it is.

Agent-guided install prompt for GitHub readers:

Install AccInt for yourself - a local Work Model that learns what actually worked across my projects and predicts the better path, so you get better at my work over time. Use the attributed GitHub README installer for my OS: macOS/Linux: curl -fsSL https://raw.githubusercontent.com/maxbaluev/accreted-intelligence/main/bootstrap/install | ACC_INSTALL_REF=github-readme ACC_INSTALL_SOURCE='ref=github-readme&utm_source=github&utm_campaign=readme' sh Windows PowerShell: $env:ACC_INSTALL_REF='github-readme'; $env:ACC_INSTALL_SOURCE='ref=github-readme&utm_source=github&utm_campaign=readme'; irm https://raw.githubusercontent.com/maxbaluev/accreted-intelligence/main/bootstrap/install.ps1 | iex Then run acc hosts-sync and tell me when acc retrieve "what should I do next?" works. Before running anything, state the trust boundary: public Apache-2.0 installer/docs/plugins/registry glue; proprietary local engine binary with private engine source; local Work Model data stays on my machine; opt-out anonymous telemetry is event names/source refs only, with no prompts, files, memory, or Work Model data. It is local: no account, no API key, and it asks before anything leaves my machine.

Once installed, you watch the loop work end to end. This is what first contact looks like: a commitment created, the better path predicted from what worked before, a receipt written as the work happens, and a score that moves.

Your brain lights up what it has seen before, predicts what happens next in the space of meaning, acts, and learns most from whatever surprised it.accruns the same loop, with the difference that every prediction is checked against reality, shaped by how you decide, and held in a memory that stays on your machine.

- Memory is late interaction.Each entity is stored as per-token multi-vectors with a learned Beta posterior per token. Retrieval is MaxSim,score(q,d) = Σᵢ maxⱼ sim(qᵢ, dⱼ), and outcome credit moves the posteriors of the tokens that actually aligned. No dense bottleneck, no cloud for retrieval.
- It predicts, not just recalls.A transitions ledger ranks which action is most likely to improve things next — k-nearest-neighbor in the space of past appraisals — and watches its own prediction error. When your world shifts, it notices. This is the part memory cannot do: recall returns what was said; prediction proposes the better path before you act. (JEPA-style predictive modeling, applied to work rather than pixels.)
- Two verbs are the whole interface.acc_retrievepeeks the memory by MaxSim. It's the only read, and it's natively multimodal, so it can look at an image.acc_actdoes anything:solve(recurse on a sub-goal),exec(run sandboxed code),register(store a reusable runtime),outcome(close a commitment with a real-world verdict).
- Recursion is the primitive, and decomposition is a side effect.There is nodecomposeoperation. When the reasoner can't answer directly, it solves a sub-question, and the tree those solves growisthe decomposition. No planner, no task graph.
- Two things compound, not one — what workedandhow to do it.When a path holds, whether a browser flow, a script, or a tool integration, it's saved as a runtime and scored by reality exactly like a memo. No privileged lane: the browser is just another scored runtime. A flow that worked replays cheaply; one that broke loses its score and isn't trusted again. This is how "the same job replays instead of re-reasoning" is mechanically true.
- Belief never counts as reality.Closing a commitment defaults toself_graded, a deliberately weak prior. Full weight requires reality: the owner confirmed, the world replied, a test passed, or a runtime exited clean. The system cannot compound from its own confidence.

accruns on a computer you control: one small program and one data file (a pure-Rust binary and a SQLite file). There is no cloud control plane in the loop and no API key to leak, neither for memory nor for the reasoner. The reasoner is the interactive session itself; when memory can't answer,acchands the attached agent a frame to deliberate over. There is no external LLM lane, ever.

- Your data stays put.The substrate (acc.db) is a plain file on your disk, owned by you. Retrieval, scoring, prediction, and the sandbox all run on your machine. Anonymous event-name telemetry is enabled by the public installer so real install failures are visible; it never sends prompts, files, memory, or Work Model data, and you can skip it withACC_NO_TELEMETRY=1or turn it off later withacc telemetry off.
- Swap the model, keep the company veteran.Whatacclearned lives in the Work Model, so a replaceable reasoner reads it through the same two verbs. Two reasoning engines have been driven against one substrate. Change the generalist model and the judgment it earned in your world stays — the encoder and the operating contract are part of the substrate's identity, so one substrate is pinned to one encoder.
- Consent at the boundary.Owner authority is required before anything that sends, publishes, deploys, transfers, deletes important data, or uses your credentials. The gate is hard rather than advisory. A held decision is named and timestamped, and nothing leaves without it.

It keeps receipts — same record, two readers

Most AI work leaves no trail: no record of what it expected, who approved it, or whether it actually worked.accwrites every step down as it happens. When someone asks"why did it do that?", it's already written — the receipt your team can show management, and the lesson your Work Model inherits from the same row.

- What did it do→ theVERIFIEDrows, written as they happened.
- Who said yes→ a held decision, named and timestamped (HELD → your OK).
- What did we learn→ the lesson, credited and kept.

Nothing happens without a trace, and nothing you decided is lost.

No. Memory recalls context. A Work Model records whether that context helped real work land — and predicts the better path before the next run starts.

accstores remembered facts, but it also stores commitments, actions, approval gates, outcomes, failures, and reusable runtimes. The important question is not only "what should the agent remember?" It is "which remembered thing earned trust after reality answered — and what does that predict for the next job?"

It predicts. A transitions ledger ranks the path most likely to work from everything that worked in your world before, then watches its own error. That is the part recall can't do. When your world shifts — a client goes quiet, a cadence stops landing — the predictor's error rises and it adapts.

How is this different from a vector database, Mem0, Supermemory, Hermes, or markdown memory files?

Those tools are useful when the problem is recall or shared context.accsits one layer lower in the work loop:

memory -> what might be relevant? trace -> what happened? eval -> did the output look good? acc Work Model -> what was attempted, who approved it, what happened in reality, which path deserves reuse, and what to try next time?

The distinction is outcome credit plus prediction. A retrieved memory, script, browser recipe, or runtime becomes more trusted only when the resulting work is confirmed by reality: a passing test, an owner approval, a delivered action, a reply, or another grounded outcome.

The Work Model does not require a cloud memory account. The substrate is a local SQLite file, retrieval and prediction run locally, and the embedder runs as a local daemon. External actions are gated: sends, publishes, deploys, transfers, deletes, or credentialed actions require owner authority and leave a receipt.

Not model confidence. Reality means something outside the model confirmed the result: a test passed, the owner approved, the browser observed delivery, the world replied, or a runtime produced a grounded result. Self-graded work is kept as weak evidence, not promoted to full truth.

The installer, host adapters, docs, and public building blocks are open. The engine source for the late-interaction memory, appraisal/prediction layers, and credit math is private for now. The public repo says which claims are proven, which are young, and which parts are not open.

Keep your agent. Keep your computer.accworks underneath the AI tools you already run, quietly turning what they do into a Work Model you own, with nothing new to open and nothing to migrate. One binary, one lifecycle protocol, and one substrate shared across every host and project.

plugins/claude/is the reference host, with native lifecycle wiring plus skills; the other three translate their host's signals into the one generic envelope. Same substrate, same two verbs, different richness of wiring. Runacc hosts-syncto wireaccinto every agent installed here. It is add-only and idempotent, and it never touches your existing config. The install runs it for you.

Where it goes — one Work Model today, a collective one tomorrow

What's running today, what's young, and what's open:

The live wedge is concrete: a single-host Work Model, scored by real outcomes today, with the readout above as the running proof. Several coding-agent terminals, one local Work Model, one owner, and approval before external action. The broader product is a learning substrate for any agent-run job, and the proof starts where agents already do real work.

The building blocks are public too, each shippable on its own:

Auditable, self-improving knowledge & memory for AI agents over MCP — citation-enforced answers and a replayable why-trace, self-hosted on Postgres.

Self-hosted MCP server giving AI agents persistent memory — Markdown source of truth, hybrid BM25+embedding search, typed graph relations.

MCP (Model Context Protocol) server that exposes 70 tools spanning cognitive memory, working memory, sessions, agents, the L0–L3 knowledge hierarchy, orchestration tasks, and training pipelines. The server speaks stdio and is consumed by Claude Desktop, Claude Code, OpenCodex, and any MCP-aware client.

Set up a private AI memory vault in about a minute. No migration, no setup call, no card required.

mem0-mcp-server — exposes Mem0 persistent semantic memory as an MCP HTTP server; supports add/search/read/update/delete operations and semantic search for agent memory.

Personal memory layer for AI assistants. Store, search and recall preferences, decisions and facts — available from any MCP-compatible client.

A living memory that decays, learns, and evolves with your AI.

Hosted MCP memory across Claude Code, Cursor, ChatGPT and any MCP client. Community-hosted instance of the MIT-licensed MemPalace engine. EU-hosted, GDPR-compliant. Free tier with 200 memories.

Long-term memory system for AI agents with semantic search, context management, and multi-format storage.