v2.5.1 — now on GHCR 🐳 — Live

Claude That
Actually Remembers.

Every session you start with Claude is a blank slate. You re-explain your stack, your decisions, your conventions — every time. DepthFusion ends that. Persistent, structured memory across every session, forever.

Install DepthFusion
0tests passing
0%Cat-D continuity score
0%token efficiency gain
0MCP tools
The hidden cost

Every new session, Claude starts completely blind.

You've spent weeks developing your architecture, making decisions, establishing patterns. Then Claude's context compacts — and it's all gone.

🧱

You re-explain everything, every time

Stack, conventions, past decisions, ongoing work. Burned tokens before you even start. The average developer re-types 300–500 tokens of context per session.

💸

40% of your context window is noise

Without memory, Claude fills your context with re-established facts instead of actual work. You're paying for repetition, not progress.

🔁

Claude contradicts its own past answers

Suggest Redis one session, Postgres the next. Without memory, every answer is context-local — there's no way for Claude to know it's reversing a decision you already made.

🫥

Context compaction erases your momentum

Hit the token limit and Claude forgets the last 3 hours of architectural context. You're back to square one, re-loading mental model into the prompt manually.

Side by side

What starting a session actually looks like.

Same question. Same developer. Two very different experiences.

✗ Without DepthFusion
G
What auth approach are we using for this project?
I don't have context about your project's authentication setup. Could you share the relevant code or describe what you're working with?
G
We decided on JWT last week. Session-scoped, 15 min expiry...
Got it. So for JWT with 15 min expiry you'd typically... [re-explaining basics you already know]
✓ With DepthFusion
G
What auth approach are we using for this project?
JWT, session-scoped with 15 min expiry — decided in S-23. Refresh logic lives in publisher.ts:callMcpTool(). You noted in that session the session-per-call pattern was intentional to avoid state leakage.
G
Perfect. Extend that pattern to the new endpoint.
On it — adapting the same session-scoped model, keeping the 15 min TTL and the initialize-per-call flow.
Interactive Demo

Watch DepthFusion in Action

Step through three real scenarios. Left panel shows what you and Claude experience — right panel shows the technical pipeline firing in sync.

User Experience Layer
Step 0 / 4
Technical Pipeline
Token intelligence

Stop paying for context you already know.

Every token in your context window either moves your project forward or wastes money re-establishing facts. DepthFusion surfaces only what's relevant — pre-scored, pre-filtered, pre-deduplicated.

Without DepthFusion
Re-explanation (~40%)
Stale/contradictory (~15%)
Actual work (~45%)
With DepthFusion
Smart context injection (~10%)
Actual work (~85–90%)
Persistent knowledge graph

Memory that grows with every session.

Without DepthFusion, sessions are isolated islands. With it, every session builds on the last — decisions connect to outcomes, incidents link to fixes, facts reinforce each other.

Decision memories
Semantic facts
Operational patterns
Episodic incidents
What you get

Everything Claude needs to think continuously.

Seven typed memory objects, eight scoring dimensions, autonomous consolidation, and contradiction detection — all working silently in the background.

🧠

Hybrid Recall Pipeline

BM25 keyword match feeds into Haiku reranking for semantic scoring. Best-of-both precision without embedding infra overhead.

BM25 → Haiku reranker → ContextRouter
⚖️

CognitiveScorer

8-dimensional weighted scoring: recency, relevance, confidence, frequency, uniqueness, coherence, actionability, and source trust.

max(confidence) over mean — per S-70 design
🗜️

SessionCompressor

When context fills, the compressor extracts decisions, open tasks, and code references — then writes them to the store as typed memory entries.

Triggered at 85% context threshold
🌐

Cross-Project Index

Query patterns, decisions, and code references across all your registered projects. Knowledge doesn't silo — it flows where you need it.

Scoped queries with project filters
🔍

ContradictionEngine

Detects when new memories conflict with established facts. Flags contradictions before they reach Claude's context — your architecture stays consistent.

Confidence-weighted suppression
📊

EventLog + GraphTraversal

Every recall, publish, and feedback event is logged. Graph traversal lets you follow memory chains — discover how a decision influenced downstream work.

SQLite-backed, local-first
By the numbers

Real performance gains, measurable from day one.

94%
Cat-D continuity score
40%
token efficiency gain
<1.5s
full recall pipeline
35
MCP tools available

"Instead of spending the first 10 minutes of every session re-loading context, Claude already knows where we left off — the architecture, the decisions, the gotchas. It's like the difference between a contractor who reads the brief and one who was there from day one."

The experience of working with DepthFusion active, across a multi-week engineering project.

How it works

Zero configuration. Automatic capture.

DepthFusion hooks into Claude Code's lifecycle events. You don't change how you work — it just starts remembering.

1

Captures as you work

PostCompact and Stop hooks automatically extract decisions, patterns, and facts from every session. The LLM extractor achieves ≥80% precision — not noisy, not lossy.

2

Scores and deduplicates

The CognitiveScorer ranks every memory block across 8 dimensions. The ContradictionEngine suppresses conflicting facts. Only relevant, accurate context reaches Claude.

3

Surfaces it automatically

At the start of each session, DepthFusion runs hybrid BM25 + Haiku recall and injects the top-K scored blocks. Claude starts every session already up to speed.

Live recall simulation

Watch memory surface in real time.

When you open a new session, DepthFusion scans your knowledge graph, scores candidates, and surfaces the right memories — all in under 1.5 seconds.

Waiting for query...
Community

Join the conversation.

DepthFusion is open source and community-driven. Ask questions, share discoveries, and help shape what gets built next.

Ready to give Claude a memory?

One MCP server. Zero cloud dependencies. Works with every Claude Code session from the moment you install it.

🐍 pip (local)

pip install depthfusion
python -m depthfusion.install.install

Local process, no GPU, no Docker required

🐳 Docker / VPS

docker pull ghcr.io/digittal-ai/depthfusion:latest

Full REST API on port 7300, persistent volume, GPU optional

View Documentation →