bellamem 0.2.2

A persistent, structured memory for LLM coding agents. Solves the context window problem — retrieves decisive context for each edit under a small token budget, measurably beating flat recency, LLM compact, and classic RAG.

RAG retrieves documents. Agents need to retrieve beliefs.

Bella is the visual brand; the Python package and CLI remain bellamem. pipx install bellamem, then bellamem save.

---

The problem

Your AI agent is like a brilliant intern with amnesia. Every morning you re-explain the project. Every afternoon it re-introduces the bug you fixed yesterday. It smiles and nods and produces confidently wrong output. Fluent, agreeable, and fundamentally untrustworthy.

Loses continuity. Today's session suggests the exact bandaid you rejected yesterday. Yesterday doesn't exist.

Hits the wall. /compact and /clear evaporate the specifics — rejected approaches, causal chains, small invariants that took ten messages to earn.

Confabulates with conviction. When it forgets, it doesn't ask to be reminded. It re-asserts the wrong approach with full confidence, and you're the one who has to catch it every time.

One root cause: agents only have working memory. The context window is the memory. When a turn falls out, it's gone — and the agent carries on unaware, filling the gap with plausible defaults. You can't guardrail a system that doesn't own its own beliefs.

Bella is the long-term memory layer. It extracts the structure of every conversation — decisions, rejected approaches, causes, self-observations — into a belief hypergraph that survives /clear, new sessions, and new days. When tomorrow's session asks about the flaky test, it loads what yesterday decided, what yesterday rejected, and why.

Measured across 15 real Claude Code projects on a developer's machine, Bella compresses the conversation portion of the context window by median 17.6× (range 3.6×–90×). The biggest sample (132k raw tokens) hit 90× — see the production curve.

---

Before and after

A real debugging session — twenty turns of dead-ends, side-questions, acknowledgments, and the actual fix at the end. Left column is what the context window holds. Right column is what Bella extracts alongside it — and what survives after /clear.

Flat session — what the context window holds	Bella hypergraph — what survives
user: test flaked again, 3rd time assistant: which one — auth or sync? user: sync_external_api, like last time assistant: ok, looking at retry.py assistant: timeout hardcoded to 2s. bump to 5s? user: bandaid, don't paper over it user: we need to know WHY it times out assistant: fair. tracing retry logic assistant: sync retries with 200ms backoff assistant: huh, only one retry — that's tight user: CI load? Mondays spike hard assistant: checking metrics dashboard... assistant: yes — CI hits the rate-limiter assistant: first retry hits same limit, fails assistant: 2s window exceeded before retry user: real fix: retry jitter, not timeout assistant: agreed. patching: backoff ± jitter user: good assistant: add latency metrics too? user: not now, focus on the fix	[retry-jitter is the fix] m=0.74 v=2 (user + assistant) ⇒ cause chain CI load → rate-limiter → first retry → 2s window exceeded before retry ⊥ rejected "bump timeout from 2s to 5s" (user: "bandaid, not a fix") __self__ observation "I reach for timeout bumps when retry semantics are the real problem"
~220 tokens · 20 turns · ordered by time · dies at /clear	~50 tokens · 4 beliefs · ordered by evidence mass · persists

Same information content, different geometry. The left column lets an agent reconstruct what was said. The right column lets it reconstruct what was decided, what was rejected, and what caused what — in far fewer tokens, and across the session boundaries where the left column can't go.

And the four items on the right are exactly the ones the agent would otherwise forget, re-suggest, or confabulate about tomorrow: a ratified decision (mass earned from two voices), a causal chain (the why), a dispute (the rejected bandaid, which Bella's edit guard will block if the agent tries it again), and a self-observation about its own reasoning pattern.

The ontology

Every concept in Bella is classified along two orthogonal axes. The axes aren't decorative — they drive retrieval, decay, and dispute handling differently. A normative decision about testing style doesn't decay the way an ephemeral observation about yesterday's build does.

Class — where does it come from and how is it used?

class	shape	meaning
invariant	hexagon	time-invariant principles and facts — they don't decay
decision	diamond	a commitment that constrains future action — revisable
observation	ellipse	a single empirical record — a snapshot
ephemeral	rounded square	pinned for short-term use, has an open→consumed/retracted/stale lifecycle

Nature — what kind of claim is this?

nature	color	meaning
metaphysical	amber	what the system IS — self-model, architectural facts
normative	blue	what we commit to — rules, preferences, policies
factual	green	measurable, checkable facts about the world

The 12 cells are a full typology — invariant × metaphysical is the deepest bedrock ("the graph is append-only"), ephemeral × factual is throwaway ("yesterday's CI was red"). Mass accumulates on every cell the same way (Jaynes log-odds from each ratifying voice), but the retrieval guard, the decay curve, and the dispute resolution read each cell differently.

You can see the full ontology rendered live on your own graph:

python -m bellamem.proto viz              # .graph/v02.html (interactive)

Shape encodes class, color encodes nature, node size encodes mass, turn-hubs make the hypergraph structure visible as rosettes.

---

Install

pipx is the recommended path — a single global bellamem command, no .venv to remember, no PATH surgery:

pipx install bellamem
# or, from a local clone:
git clone https://github.com/immartian/bellamem
pipx install -e ./bellamem                  # editable install, still global

Per-project venv also works:

cd your-project
python3 -m venv .venv
.venv/bin/pip install bellamem

Optional extras:

pipx inject bellamem 'sentence-transformers>=2.2'   # local embeddings
pipx inject bellamem 'openai>=1.0'                  # OpenAI embeddings + LLM EW
# or with pip:
pip install 'bellamem[st]'      # sentence-transformers
pip install 'bellamem[openai]'  # OpenAI
pip install 'bellamem[all]'     # both

Copy .env.example → .env in your project and fill in the backends you enabled. .env is gitignored.

Requirements: Python 3.10+. Git (Bella scopes per-project state via the git repo root). No other system dependencies.

---

Quickstart

Three retrieval modes — one for each question you actually ask about your memory. Most workflows live in these three commands:

Command	Question
bellamem expand "X"	What do we believe about X, ranked by importance?
bellamem surprises	What just changed — what mattered?
bellamem replay [X]	What did we say — in what order?

Plus utility commands for ingest, audit, render, prune, and bench:

# Ingest Claude Code sessions for the current project.
# Auto-runs R3 consolidation (merges near-duplicates) on new claims.
bellamem save

# Three retrieval modes — same memory, different questions:
bellamem expand "what did we decide about persistence"
bellamem surprises                                      # top jumps, sign flips, disputes
bellamem replay                                         # narrative timeline
bellamem replay "ad-hoc bandaid pattern"                # focused narrative

# The pre-edit pack: no recency, surfaces invariants + disputes + causes
bellamem before-edit "should I wrap this in try/except" --entity embed.py

# Health report: bandaid piles, duplicates, garbage field names, mass limbo
bellamem audit

# Render the graph as a picture (needs the [viz] extra or graphviz CLI)
bellamem render --out graph.svg                           # whole forest
bellamem render --out disputes.svg --disputes-only        # just ⊥ edges
bellamem render --out auth.svg --focus "auth tokens"      # subgraph around a focus

# Forget orphan leaves that never earned their place (dry run by default)
bellamem prune                        # preview candidates
bellamem prune --apply                # actually remove them

# Empirically compare context strategies (flat, compact, RAG, Bella)
bellamem bench

Every command except save, emerge, prune --apply, and scrub is read-only.

---

Use with Claude Code

The flow that lets you keep working past the context window without losing the thread is packaged into four slash commands.

Bella is built for — and tested daily against — Claude Code. Other coding agents (Codex, Cursor, Windsurf, OpenClaw) will need their own adapter; see #1 for the lifecycle hooks we're tracking.

Install the slash command — once, globally

bellamem install-commands           # writes ~/.claude/commands/bellamem.md

/bellamem now works in every Claude Code project on your machine. Per-project install (--project) is also supported if you want to commit the slash command into a specific repo.

The commands

Command	What it does
/bellamem or /bellamem resume	Working-memory replay tail + long-term expand pack + top surprises. Run at session start.
/bellamem save	Ingest the current session (auto-consolidates), run audit, report top new surprises. Run before /clear or at end of day.
/bellamem recall <topic>	Mass-ranked beliefs about a topic, disputes included. Mid-session lookup.
/bellamem why <topic>	Pre-edit pack: invariants, disputes, causes, entity bridges. Run before a risky change.
/bellamem replay / /bellamem audit	Raw CLI output when you want to look at it directly.

The save → clear → resume flow

/bellamem save     ← captures this session into the graph
/clear             ← wipe the context window (Claude Code built-in)
/bellamem resume   ← fresh assistant reconstructs where you were

On a well-tuned project, /bellamem resume comes back in ~30k tokens and contains enough to pick up the next decision without re-asking questions already answered. If it's much larger, run bellamem emerge to consolidate near-duplicates.

The edit guard (v0.0.4)

Install bellamem-guard as a Claude Code PreToolUse hook and an advisory pack (invariants + disputes + causes for the focus) is injected automatically before every Edit / Write / MultiEdit call — no manual invocation needed. The guard exit-2s when the edit re-suggests a rejected approach (a ⊥ dispute), refusing the tool call at the boundary.

Hook registration (once per project) in .claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      { "matcher": "Edit|Write|MultiEdit", "hooks": [{ "type": "command", "command": "bellamem-guard" }] }
    ]
  }
}

Where your data lives

~/.claude/commands/
  bellamem.md            installed once (global slash command)

<your-project>/
  .claude/settings.json  PreToolUse hook registration (optional)
  .graph/
    default.json          belief graph (gitignored by default)
    default.emb.bin       belief embeddings, v3 binary sidecar
    embed_cache.json      embedding cache (pruned to live beliefs on save)
    llm_ew_cache.json     LLM EW cache (if BELLAMEM_EW=hybrid)
  .env                    your API keys + embedder choice (never commit)

.graph/ is gitignored by default.

---

Bella vs /compact

Both compress a long session. The difference is load-bearing:

	/compact	Bella
Output	One narrative summary (~2000 tokens)	Queryable belief graph (~3k per retrieval)
Shape	Prose	Beliefs + typed edges (→, ⊥, ⇒) + mass + voices + sources
Usage	Replaces history; summary becomes new context	Load on demand per turn; three retrieval modes
Preserves	Broad topics, major decisions, flow	Paraphrased decisions, rejected approaches, cause-effect chains, self-observations, line numbers
Loses	Identifiers, ⊥ corrections, causal structure	Tool outputs, file contents, conversational texture
Cross-session	None — dies with the session	Full — graph persists, next session inherits it

On our 13-item bench, compact scored 8% LLM-judge; Bella's expand scored 92%. Narrative summaries preserve themes; structured retrieval preserves decisions. The two are complementary: /compact keeps the feel inside one session; Bella keeps the decisions across sessions.

---

See the graph

A real Bella belief hypergraph — roughly 1,800 beliefs from a month of Claude Code sessions on Bella itself, across eight topical fields. Typed edges (→ support, ⊥ rejected approach, ⇒ cause) and mass-weighted nodes are the structure that flat context can't preserve.

Click the image to play the .webm.

---

Empirical results

Latest measurement: benchmarks/v0.0.4rc1.md (2026-04-10, budget = 1200 tokens, LLM judge enabled, 13-item hand-labeled corpus, 1834-belief forest).

metric               flat_tail      compact     rag_topk       expand  before_edit
----------------------------------------------------------------------------------
exact hit rate            15 %          0 %         15 %         69 %         46 %
embed hit rate            23 %         31 %         31 %         85 %         77 %
llm judge rate             0 %          8 %         31 %         92 %         69 %
avg tokens used           1200          602         1161         1143          964

flat_tail (0%) < compact (8%) < rag_topk (31%) < before_edit (69%) < expand (92%).

Headline story — compare to v0.0.2: as the forest grew from the v0.0.2 dogfood snapshot to 1834 beliefs, rag_topk collapsed from 85% → 31% LLM judge (cosine top-k pulls up more plausible-looking-but-wrong neighbors in a larger forest), while expand held at 92%. The gap from expand to the next-best contender widened from 15pp to 61pp. Structured mass-weighted retrieval scales with forest size; cosine top-k doesn't. The retrieval code path (core/expand.py, core/bella.py) is unchanged between v0.0.2 and v0.0.4rc1 — every delta is a property of forest growth, not algorithm changes.

See benchmarks/README.md for the versioning convention and when to re-run.

Compression at scale — 15 real Claude Code projects

The bench above answers "is expand accurate?" The next question is "how many tokens does Bella actually save?" For that, the docs/scenarios.md harness measures real Claude Code session transcripts sampled from 15 different projects on a developer's machine — news monorepos, IRB documents, refactoring sessions, agent prototypes, marketing work — at a fixed expand budget of 1500 tokens.

	tokens
raw conversation range	274 → 132,399
compression ratio range	3.6× → 90×
median ratio	17.6×

Visually:

The pattern is unambiguous: expand honors whatever budget the caller passes, regardless of how big the raw transcript got. The horizontal line in the chart is the budget I picked for these measurements (1500) — at budget=3000 every ratio would halve, at budget=500 every ratio would triple. The actual claim is that the ratio diverges with raw size at any budget you choose. Doubling raw doesn't double expand; it doubles the ratio. The biggest sample (a multi-day news monorepo session at 132k conversation tokens) hit 90×.

A second chart in the same doc shows the smaller-scale linear regime where expand grows with raw — that's where the synthetic break-even math gives the rule of thumb "don't bother with Bella for conversations under ~200 tokens; the per-belief overhead dominates." Above that, Bella pays off, and the longer the session, the more it saves.

Sources are anonymised; only aggregate metrics are pinned. The measurement script lives in docs/scenarios.py for anyone who wants to reproduce on their own ~/.claude/projects/.

---

Limitations

Bella lives alongside the agent, not inside it. That boundary is load-bearing but not permanent — tracked as #1: Context lifecycle dependencies, blocked on upstream anthropics/claude-code#47023 (exposing compact/session lifecycle hooks). Today what we have is advisory:

• No direct context-window control. Bella can't swap active tokens, evict irrelevant context, or replace the window wholesale. The agent still controls what it attends to; Bella can only offer packs the agent can choose to read.
• /compact stays LLM-driven. Claude Code's native /compact writes a narrative summary via an LLM call. A PreCompact hook that lets Bella substitute a graph-backed compaction would unlock most of the remaining wins — and that hook surface does not exist in Claude Code today. We can't intercept it from outside.
• Save/clear/resume is a manual pattern. You run /bellamem save → /clear → /bellamem resume yourself. It works, but it's a human-in-the-loop ritual, not an autonomous context manager.
• The edit guard is a tool-call boundary, not a semantic gate. bellamem-guard injects an advisory pack before every edit and exit-2s on a dispute re-suggestion, but it sees tool-call text, not model intent. An agent that ignores the advice can still try the edit; the block is at the boundary, not deeper in the model.
• One adapter at a time. Claude Code works today. Codex and others need their own turn-pair reaction classifier and source stamper.

The common thread: every limitation above is about how much of the agent's context lifecycle we can observe and influence from outside. With deeper hooks — or a coding agent that exposed its context as a first-class API — a graph memory like Bella could drive the compaction cycle itself instead of being handed the leftovers. We expect the upside of real context-window control to be substantial. These aren't permanent trade-offs — they're tracked in #1, resolvable either upstream (via claude-code#47023) or downstream (transcript-watcher fallback). For now, the honest frame is: Bella is the memory layer; the agent is still the window manager.

---

Status

v0.1.0 — alpha, dogfooded on its own construction. Bella was built in Claude Code sessions that were themselves ingested into the Bella being built. When the assistant drifted into an ad-hoc bandaid pattern during development, the user's correction landed in the graph as the highest-surprise belief of the session. That kind of self-observation is the point.

Since v0.0.2:

• v0.0.3 — per-project .graph/, automatic R3 consolidation on ingest, source grounding + narrative replay, structural pruning, bellamem save default-to-current-session with incremental ingest, and embed-cache prune bounded to live beliefs.
• v0.0.4rc1 — storage split: belief embeddings moved out of default.json into a default.emb.bin sidecar (v3 format), cutting non-vector operations' load time from ~2s to ~500ms. bellamem-guard PreToolUse hook ships: advisory pack before every edit, exit-2 block on dispute re-suggestions. Embedder batching reduces save latency.
• v0.1.0 — log-odds decay gated on BELLAMEM_DECAY=on: on every save, non-exempt beliefs fade exponentially toward the 0.5 prior at a 30-day half-life (reserved fields, mass_floor pins, ⊥ disputes, and ⇒ causes are exempt). New bellamem decay subcommand for dry-run preview + --apply. v3 → v4 snapshot format adds a decayed_at header. See the "Decay and reinforcement — the steady state" section of THEORY.md for the collision math.
• v0.1.1 (planned) — decay on by default after dogfood validates the steady state, Three.js 3D viz with temporal replay, and graph-backed compaction when the hook surface allows.

See CHANGELOG.md for details.

---

Architecture

bellamem.core (Jaynes accumulation, expand, audit, prune, replay) never imports from bellamem.adapters (Claude Code reader, LLM EW, turn-pair classifier). Core is domain-agnostic; adapters are where domain knowledge lives. Seven mutation operations, pluggable embedders, split snapshot storage.

Full file tree and invariants: ARCHITECTURE.md.

---

Theory

The formal calculus — six rules, invariants, Bayesian grounding, domain-agnostic case studies — lives in bella/:

• SPEC.md — the six rules, formal definitions
• VISION.md — theoretical grounding (Jaynes, Gödel, self-reference)
• EXAMPLES.md — case studies (H. pylori, continental drift, …)
• MEMORY.md — how BELLA maps to LLM agent memory

THEORY.md covers the implementation side: how bellamem maps the BELLA spec to Python — thresholds, data structures, a worked flaky-test example with before/after diagrams, and the decay equilibrium math.

---

Contributing

See CONTRIBUTING.md. Short version:

• The bench is the CI. Run bellamem bench after changes to EW, expand, or audit and report the delta in the PR.
• Add new embedders by implementing the Embedder protocol in core/embed.py.
• Add new EW logic in adapters/, never in core/.
• Every PR that touches retrieval should include a bench item demonstrating the failure mode it fixes.
• Dogfood changes against Bella's own snapshot before shipping. Unit tests prove code runs; running Bella against its own graph proves the feature is useful.

---

License

AGPL-3.0-or-later. See LICENSE.