TokenMizer

Your AI forgets why. TokenMizer remembers.

An OpenAI-compatible proxy that builds a knowledge graph of your session — decisions, files, errors, goals — and replays it when the
context window runs out. Not a summary: a queryable graph that knows "we switched from MongoDB to PostgreSQL, and here is why."

_{One line to adopt · works with Claude, GPT, Gemini, Grok, DeepSeek, Mistral, Cohere, Ollama · MIT}

Quick start · See it · Claude Code & MCP · Benchmarks · Architecture · Roadmap

TokenMizer demo: a 40-turn session checkpointed at 87% context, resumed next day in 233 tokens

_{Real run: 25-node graph, checkpoint ckpt_21a0959c3ddf, 233-token resume. Regenerate with python scripts/gen_demo_gif.py.}

--- ## The problem Every AI session has a context limit. When you hit it, the model forgets every decision and every rationale built over hours of work, and you spend the first ten minutes of the next session re-explaining the project. Summarising the history does not fix this. A summary tells you *what* was decided; it loses *why*, and it loses what was rejected — so the model happily re-proposes the thing you moved off three sessions ago. ## Quick start ```bash pip install "tokenmizer[anthropic,cache]" export TOKENMIZER_ANTHROPIC_API_KEY=sk-ant-... tokenmizer serve ``` Then change one line in your client: ```python from openai import OpenAI client = OpenAI( api_key="your-key", base_url="http://localhost:8000/v1", # only this changes ) resp = client.chat.completions.create( model="claude-sonnet-4-6", messages=[{"role": "user", "content": "Continue where we left off"}], extra_body={"session_id": "my-project"}, # optional, enables memory ) ``` Everything else is unchanged: same request shape, same response shape, plus a `tokenmizer` block reporting what was saved. Open and the session is already there.

Windows, Ollama, Docker, and the full step-by-step

**Windows (PowerShell)** ```powershell $env:TOKENMIZER_ANTHROPIC_API_KEY = "sk-ant-..." # this session setx TOKENMIZER_ANTHROPIC_API_KEY "sk-ant-..." # persistent ``` **No API key?** Ollama runs locally and free: ```bash ollama pull llama3 pip install tokenmizer # then set `provider: ollama` in tokenmizer.yaml ``` **Docker** ```bash docker compose up -d ``` Full installation notes, every provider's environment variable, and the configuration reference are in [**docs/configuration.md**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/configuration.md) and [**docs/deployment.md**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/deployment.md).

## See it Nothing here is a mockup. Every screenshot and the demo below are the shipped UI rendering a session from the labelled corpus in `benchmarks/eval/corpus`. ### The session graph is a page you drive, not a picture

The graph explorer in use: the radial map, a node selected to show its type, confidence, first-seen date and every relation, the force layout with communities and unconnected facts, a node type filtered out, the timeline, and the light theme

_{Radial → select a node → force → filter a type → timeline → light. Regenerate with python scripts/gen_graph_demo.py.}

Select a node and you get its type, status, community, importance, confidence, when it was first seen and every relation it carries — the provenance behind a fact, not just the fact. Filter a type or a community and the counts move with it. Three layouts, a light theme and a PNG export are one click each. It is **one self-contained HTML file with no external requests**, so it opens offline, works from a `file://` URL, and can be sent to someone who has never installed TokenMizer. ### The dashboard tells you what it actually knows

TokenMizer dashboard: savings, sessions, the live resume block and an embedded radial session graph

Your sessions, the live resume block each one would inject right now, and its graph — not an example of one. The health pill reads `/health`, which reports `degraded` with the counters behind it when a write has failed, rather than saying `ok` whatever happened. ### The session graph, grouped the way you would group it

The session graph as a radial map: one arc per node type, named on the ring, with relations drawn as chords through the middle and a panel of session numbers, communities and hotspots

Each node type gets its own arc of the circle, named on the ring, and relations are drawn as chords bowed through the middle — so the shape of the session is readable before you read a single label, and which type a node is comes from *where it sits*, not from telling two hues apart. The palette is checked with a validator, not by eye: every adjacent pair clears the colour-blind separation floor, and the four types that carry no meaning of their own share one neutral grey. The panel beside it counts what the session knows and what is missing — open issues, decisions changed, history gaps, unconnected nodes — then lists the detected communities, the hotspots everything hangs off, and which kinds of node point at which. Filter by type or community, search, click a node for the supersession chain behind it. ### The same session as a story

Timeline view: decisions, tasks, errors and files in their own lanes, with a supersession arc

Timeline mode puts each node type in its own lane, ordered by when the fact entered the session, with supersessions drawn as arcs. When a whole transcript was checkpointed in one call every node shares a timestamp, so the axis says so and falls back to the order the session stated things rather than inventing dates. ## How it works TokenMizer is a local proxy between your app and any LLM. Every request passes through a pipeline that builds a live knowledge graph, compresses inputs, caches responses, and checkpoints before the context runs out. ```mermaid flowchart LR App["Your app
_{OpenAI-compatible client}"] subgraph TM["TokenMizer :8000"] direction TB L0["L0 File intelligence"] L1["L1 Prompt compression"] L2["L2 Terse-output injection"] L4["L4 Graph memory
_{extract to window to inject}"] L3["L3 Semantic cache"] L5["L5 Provider prompt cache"] L0 --> L1 --> L2 --> L4 --> L3 --> L5 end LLM["Claude · GPT · Gemini
Grok · DeepSeek · Ollama"] DB[("SQLite
_{graph · checkpoints · ownership}")] App -->|"POST /v1/chat/completions"| TM TM --> LLM LLM -.->|response| TM TM -.->|"response + savings"| App L4 <-->|"per-row, locked"| DB ``` The graph is not a summary. It is typed nodes and edges — decisions, tasks, files, errors, goals — with a lifecycle, so a decision that gets replaced is marked superseded rather than deleted, and the transition records what triggered it. The resume block is a filtered projection of it: active decisions, open work, unresolved errors, in a few hundred tokens. Edges carry the relations a session actually has. A task that fixed a bug `FIXES` the error node and closes it. An open error `BLOCKS` the task about it. A decision `DEPENDS_ON` the package it named. That is what the communities above are detected from, and what `/why` walks. To [**Architecture**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/architecture.md) — the request sequence, the data model, and the decision lifecycle. ## What a resume looks like ``` Goal: FastAPI authentication service with JWT and PostgreSQL Working on: refresh token rotation in api/auth.py | rate limiting using slowapi Done: Implemented POST /api/auth/login | Fixed the 422 in LoginRequest | User model in api/models.py Decided: Use JWT and PostgreSQL | bcrypt for password hashing | Redis for refresh token storage Changes: 'Use moment.js' -> 'Use date-fns' - tree-shakeable, saves 230KB Files: api/auth.py, api/models.py, config.py, tests/test_auth.py Continue from: Add rate limiting to auth endpoints ``` A few hundred tokens in place of the whole conversation. The `Changes:` line is the part a summary loses — and `GET /api/graph/{session_id}/why?q=date-fns` replays the full chain with the trigger, the reason and the evidence for each hop. ### Habits, not just projects A session graph remembers what you decided *about this project*. It does not remember that you want short answers — that is true of you, not of the repository, and it has to survive starting a session somewhere else. ```yaml preferences: enabled: true # off by default; read the note before turning it on ``` With it on, a turn like "I prefer TypeScript, and keep answers brief" is remembered per principal and injected as a few lines of system prompt. `/api/preferences` (GET) shows exactly what was remembered and the exact text it injects; the same path with DELETE and `?key=...` forgets one, without a key forgets all of them. **Off by default on purpose.** The failure mode of a preference memory is not forgetting — it is remembering something that was never a preference and repeating it in every prompt you send for the rest of the year. The detector is a set of regexes and it will have false positives, which is why the endpoint above exists and why you turn this on deliberately. ### Not only coding sessions Set `domain` and the same five shapes are read in another vocabulary. Nobody in an incident channel says "Decided:", and nobody in a research log says "Fixed:" — which is why those sessions used to come back nearly empty (**macro F1 11%**, decisions and errors at **0%**). ```yaml domain: ops # coding (default) · research · ops · product ``` ``` Incident: We are seeing an incident on the pricing service, checkout is failing Mitigating: Still monitoring the replica lag | Follow-up to add a pool-size alert Done: Restarted the pool workers and rolled back to build 4471 | Traffic is recovered Decided: Root cause is connection pool exhaustion after the 14:02 deploy Symptoms: Error rate is 34 percent and p99 latency jumped to 8 seconds ``` **11% to 93%** on the labelled sessions in `benchmarks/eval/corpus_domains`, with the coding corpus unchanged — a pack's patterns run *after* the coding ones and can only add. Three hand-written sessions, so read it as "the mechanism works on sessions of this shape", not as a generalisation claim; the [benchmarks](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/benchmarks.md) say the same. ## Use it from your tools Four ways in, depending on where you work. All of them talk to the same graph, so a session checkpointed from Claude Code resumes in the CLI. ### Claude Code — plugin ``` /plugin marketplace add Shweta-Mishra-ai/tokenmizer /plugin install tokenmizer@Shweta-Mishra-ai/tokenmizer ``` Then, in any session: ``` /tokenmizer:checkpoint my-project save the session to graph memory /tokenmizer:resume my-project load it back (~300 tokens) /tokenmizer:analyze data/sales.csv digest a large file /tokenmizer:stats token savings report ``` ### Claude Desktop, Cursor, VS Code, Zed — MCP server ```json { "mcpServers": { "tokenmizer": { "command": "tokenmizer-mcp", "env": { "TOKENMIZER_URL": "http://localhost:8000" } } } } ``` | Client | Where that goes | |---|---| | Claude Desktop (macOS) | `~/Library/Application Support/Claude/claude_desktop_config.json` | | Claude Desktop (Windows) | `%APPDATA%\Claude\claude_desktop_config.json` | | Claude Code | `.mcp.json` in the project, or `~/.claude/settings.json` | | Cursor | Settings, MCP, Add server, same JSON | | VS Code / Zed | their MCP settings, same `command` and `env` | | Codex CLI | `~/.codex/config.toml` — TOML, see [docs/api.md](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/api.md) | Restart the client afterwards. **You do not need to start anything else:** `checkpoint_session`, `resume_session`, `get_graph_stats`, `why_decision` and `analyze_file` all read and write the graph directly when `tokenmizer serve` is not running, and say which path answered. Only savings need the proxy, since savings are measured on requests that pass through it. If `tokenmizer-mcp` is not on your PATH, use `"command": "python", "args": ["-m", "tokenmizer.mcp.server"]`. **Six tools:** `checkpoint_session`, `resume_session`, `get_graph_stats`, `get_savings_stats`, `analyze_file`, and `why_decision` — ask your agent *"why did we pick X?"* and it walks the supersession chain with the reason and evidence for each hop. ### Anything else — the proxy Any OpenAI-compatible client works by pointing `base_url` at `http://localhost:8000/v1`, as in the quick start above. That covers Continue.dev, Aider, LangChain, LlamaIndex, the OpenAI SDKs in every language, and `curl`. **Tool calling goes through too.** Send `tools` / `tool_choice` in the OpenAI shape and get `message.tool_calls` back, streamed or not, with `role: "tool"` results round-tripping to the model: ```python resp = client.chat.completions.create( model="claude-sonnet-4-6", messages=[{"role": "user", "content": "weather in Pune?"}], tools=[{"type": "function", "function": { "name": "get_weather", "parameters": {"type": "object", "properties": {"city": {"type": "string"}}}}}], extra_body={"session_id": "my-project"}, ) call = resp.choices[0].message.tool_calls[0] # finish_reason == "tool_calls" ``` Native for OpenAI, DeepSeek, Mistral, OpenRouter and Grok; translated for Anthropic, Ollama, Gemini and Cohere. **All nine providers**, streamed or not — each one's tool calls arrive on the stream as OpenAI-shaped deltas, whether the provider fragments them (Anthropic, Cohere) or sends them whole (Ollama, Gemini). Tool traffic is never compressed, and a tool-call turn is never cached. To [**API & CLI reference**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/api.md) — every endpoint, every command, every MCP tool. ### Inside an agent — in-process, no server For an agent loop that already owns its messages, use the memory directly. No proxy, no API key, no network; the same SQLite store the proxy uses, so an agent and the proxy can share a session. ```python from tokenmizer.agents import Memory memory = Memory("order-service") # default: the working directory name memory.add(messages) # {"role", "content"} dicts; idempotent memory.search("what are we storing orders in", top_k=5) memory.context(token_budget=400) # resume block for the system prompt memory.why("postgres") # the decision trail ``` `search` returns plain dicts (`type`, `label`, `summary`, `status`, `confidence`); `decisions()` and `errors()` list each category. Works with LangGraph, CrewAI, AutoGen or a hand-written loop the same way: call `add` with the conversation so far, put `context()` in the system prompt. ## Measured `python -m benchmarks.eval` scores extraction against a labelled corpus of 14 sessions, 6 of them real transcripts: | Category | Precision | Recall | F1 | |---|---|---|---| | Files | 100% | 100% | **100%** | | Decisions | 97% | 100% | **99%** | | Completed tasks | 98% | 98% | **98%** | | Pending tasks | 100% | 90% | **95%** | | Errors | 93% | 96% | **94%** | | | | **macro F1** | **97%** | **Precision is reported, not just recall.** An extractor that emits the whole transcript as one node scores 100% recall, which is why recall-only extraction numbers should be distrusted — including our own earlier ones. Scored separately by origin, because hand-written fixtures are easier than real transcripts and a single headline hides that: **synthetic 98%, real 91%.** Treat 91% as the number that describes real sessions. n=14 is a small sample and the same person wrote every label. **Retrieval is measured separately.** `python -m benchmarks.graph_retrieval.query_eval` scores what `query()` returns for questions phrased the way a person asks them, not in the node's own words: **recall@6 88% over 40 cases**, keyword ranking only. The eval was 13 cases until this branch, where one case flipping moved the headline by 8 points; every case is checked to be answerable from its own transcript, because an ungrounded question measures extraction and reads as a retrieval failure forever. `semantic_retrieval: auto` turns on embedding similarity when the model actually loads — the 92% figure previously quoted for it predates the enlarged eval and has not been re-measured. **Input cost is measured end to end.** `python -m benchmarks.savings.runner` replays a conversation through the real proxy against a stand-in that applies Anthropic's documented prompt-caching rules. Against a client sending its history with no caching, input cost is **61% lower at 40 turns, 87% at 150 and 93% at 300**; against a client that already caches its own history, 10% higher at 40 turns, level at 150 and 41% lower at 300. **A ten-turn session costs 21% more input**: it is under the cache minimum, and the ~50-token brevity instruction is charged on every request while the output it shortens is not measured. Details and caveats in [Benchmarks](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/benchmarks.md). **Agent sessions are measured separately**, because tool output, not prose, is most of what they send. On a real 23-request agent session the proxy sent 70% fewer tokens than the client alone, where before this release it sent none fewer, cost 20% less than a client caching its own history, and still had 97% of what the agent went on to use in the request (94% over 58 requests). That is one session and an input-only estimate, and information retained is not answer quality: see [Benchmarks](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/benchmarks.md). **Independently verified against 7 other methods.** A separate 100-session benchmark ([tokenmizer-research](https://github.com/Shweta-Mishra-ai/tokenmizer-research), a different corpus and scorer than the numbers above) ties TokenMizer 0.5.4 for first place at **60% macro F1** — level with Mem0-style (60%) and Graphiti-style (59%), ahead of GraphRAG-style (44%), MemGPT-style (35%), and every naive baseline (under 20%). **Measured on real agent traffic, not only on labelled corpora.** The same research repository runs the product on **1,467 public SWE-bench agent sessions**: SWE-agent with GPT-4, GPT-4o, Claude 3 Opus and Claude 3.5 Sonnet, and OpenHands with function calling, about 31 M tokens in total. No label is written by hand: - the files the agent edited come from its submitted patch; - its errors come from tracebacks and linter output in tool results. Sessions are split by issue. Bugs were found on one half only, and the other half (716 sessions) was scored once, before and after the fixes: | Test split, 716 sessions | Before | After | |---|---:|---:| | Runtime errors recalled | 63% | **92%** | | Runtime errors that reach the resume block | 20% | **50%** | | Edited files that reach the resume block | 51% | **67%** | | Extracted errors that match a real error (lower bound) | 16% | **31%** | | Edited-file recall | 75% | 78% | In the production condition, one message per call, runtime-error recall went from 72% to **96%**. The comparison methods recall 5–12% of those runtime errors. Full tables, confidence intervals and threats to validity: [`REPORT_real_agents.md`](https://github.com/Shweta-Mishra-ai/tokenmizer-research/blob/main/benchmarks/results/REPORT_real_agents.md). To [**Benchmarks**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/benchmarks.md) — memory quality against a plain-summary baseline, storage, and how to score your own sessions. ## Why TokenMizer and not X? **Why not just use Git history?** Git stores *what changed*, not *why you decided to change it*. You cannot ask Git "what did we decide about auth?" or "why did we switch from MySQL to PostgreSQL?" TokenMizer stores decisions with trigger, reason, and evidence — not diffs. **Why not RAG (retrieval-augmented generation)?** RAG retrieves *relevant chunks* — it does not model *decision state*. If you switched from bcrypt to Argon2 mid-session, RAG might retrieve both and confuse the model about which is current. TokenMizer tracks decision supersession explicitly: the old decision is marked `SUPERSEDED`, the new one `ACTIVE`, and the resume context only includes current state. **Why not a plain summary at the start of each session?** Summaries lose structure. You cannot query "all superseded decisions" or "what triggered the auth change" from a blob of text. Our benchmark shows graph memory preserves **89%** of labelled information against **79%** for a summary baseline — and unlike a summary, the graph is queryable, editable, and grows incrementally instead of being re-summarized every turn. See [Benchmarks](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/benchmarks.md#memory-quality--graph-vs-a-plain-summary). **Why not Mem0 or Zep?** Mem0 and Zep store *facts* ("user prefers Python"). TokenMizer stores *decisions with rationale* — the full causal chain: what was decided, what replaced it, why, what evidence triggered the change. If you need "remember my name across sessions," use Mem0. If you need "remember that we switched from PostgreSQL to SQLite because of cost, and here is the evidence," use TokenMizer. **Why not Graphiti?** Both build a temporal knowledge graph, and on the independent benchmark above they score within a point of each other. The differences are operational: TokenMizer runs on SQLite with no database to deploy, ships the graph as a self-contained HTML page you can open offline or send to someone, and is an OpenAI-compatible proxy — so adopting it is a base URL change rather than an integration. Graphiti is the better fit if you are already running Neo4j and want to query the graph in Cypher. **Why not just a longer context window?** Longer context means higher cost, slower inference, and attention dilution on long histories. TokenMizer compresses a session into a resume block averaging **161 tokens** (measured, n=3 — see [Benchmarks](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/benchmarks.md)) by extracting what actually matters, not by summarizing. ## What is not implemented Listed here rather than left to be discovered: | Setting | Status | |---|---| | `routing.*` | **Deprecated, removed next release.** Never implemented. Use `model_map` (below) to send one model name to another — that is what it was reached for. A config carrying a `routing:` block still loads and logs a deprecation warning. | | `state_backend: redis` | Accepted, never implemented; behaves as `memory` and warns at startup. Use `state_backend: sqlite`, which shares the rate limiter across workers on one host. | | `functions` / `function_call` (the deprecated OpenAI shape) | Accepted and ignored, with a server-side warning. Use `tools` / `tool_choice`, which are forwarded. | The prioritised plan for these and everything else is in [**docs/roadmap.md**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/roadmap.md), which pairs every planned item with the measurement that motivates it. ## Documentation | | | |---|---| | [**Architecture**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/architecture.md) | Request pipeline, graph data model, decision lifecycle, file intelligence | | [**Configuration**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/configuration.md) | Every setting, environment variables, precedence, providers | | [**API & CLI**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/api.md) | Endpoints, commands, MCP tools, tool calling, Claude Code integration | | [**Deployment**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/deployment.md) | Docker, multiple workers, durability, session isolation, security | | [**Benchmarks**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/benchmarks.md) | Extraction quality, memory quality, storage, running your own | | [**Roadmap**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/roadmap.md) | Measured state of every layer, and the prioritised plan | | [**Comparisons**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/docs/comparisons.md) | Running alongside other token tools | | [**Contributing**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/CONTRIBUTING.md) | Setup, layer rules, and how to improve extraction | | [**Testing**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/TESTING.md) | How to run the suite, the coverage floor, and known limits | | [**Changelog**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/CHANGELOG.md) · [**Security**](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/SECURITY.md) | Release history and how to report a vulnerability | ## Contributing ```bash git clone https://github.com/Shweta-Mishra-ai/tokenmizer cd tokenmizer pip install -e ".[dev]" pytest tests/ -q && ruff check tokenmizer/ # 2131 tests, must stay green ``` **The most valuable contribution is a session where extraction got it wrong.** The eval corpus is 14 sessions and the same person wrote every label in it — that is the honest ceiling on what the numbers above can tell you about *your* workload, and the only way past it is transcripts nobody here wrote. Label a few of your own in the format documented in [`benchmarks/eval/corpus.py`](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/benchmarks/eval/corpus.py) and open a PR, or [open an issue](https://github.com/Shweta-Mishra-ai/tokenmizer/issues) with the turn that was missed. Redact freely — the shape of the prose is what matters, not its content. [CONTRIBUTING.md](https://github.com/Shweta-Mishra-ai/tokenmizer/blob/main/CONTRIBUTING.md) covers setup, the layer rules, and how to run the eval harness. If you have a question about the project or want to discuss something, join the [Discord server](https://discord.gg/fPg9KvjRe). ## Contributors We extend our gratitude to everyone who has had a pull request merged into TokenMizer. The avatars below are the authors of those pull requests, in the order they first merged one. TokenMizer is a community-driven project, and your contributions continually improve it. The full list is on the [contributors graph](https://github.com/Shweta-Mishra-ai/tokenmizer/graphs/contributors).

## Support If TokenMizer is useful to you, please give it a [star](https://github.com/Shweta-Mishra-ai/tokenmizer). It takes a second and it genuinely helps. [Sponsorship](https://github.com/sponsors/Shweta-Mishra-ai) is open too, if you would like to support the work. Entirely optional. ## License MIT © [Shweta Mishra](https://github.com/Shweta-Mishra-ai)