nervapack 0.5.2

Privacy-first, offline knowledge graph for developers

NervaPack

NervaPack is a privacy-first, offline knowledge graph for your codebase. It solves two fundamental problems with standard Vector RAG:

• Token waste — chunk-based RAG retrieves blobs of text that may only tangentially relate to your query, bloating your context window.
• Privacy risk — sending code to cloud embedding APIs leaks your proprietary logic.

NervaPack runs 100% on your machine. It uses tree-sitter to parse your codebase into a deterministic Abstract Syntax Tree graph, then uses a local Ollama model to draw hard semantic edges between your documentation and your code. Queries traverse this graph with a K-Hop BFS, returning a hyper-targeted, token-efficient context window — no cloud required.

---

⚡ Verified Performance

91.2% Average Token Reduction — independently verified on real-world codebases.

Test Type	Tokens (Naive)	Tokens (NervaPack)	Reduction
Simple Query	10,926	101	99.1%
Medium Query	13,092	164	98.7%
Complex Query	3,290	1,102	66.5%
Average	52,037	2,459	91.2%

Cost Savings: $181-$724 per developer per year (GPT-4o to GPT-4 Turbo)

📊 View Full Benchmarks · 🧪 Messy Code Performance

Code Quality Impact: 90-99% reduction on clean code, 50-75% on legacy/messy code. Even poorly structured codebases benefit significantly.

---

Why NervaPack vs. standard Vector RAG

	Standard Vector RAG	NervaPack
Parsing	Arbitrary text chunks	Deterministic AST nodes (class, function, import)
Retrieval	Nearest-neighbor blob	K-Hop BFS on a structural graph
Doc ↔ Code links	None	Hard EXPLAINS edges drawn by local LLM
Privacy	Cloud embeddings	100% local (ChromaDB + Ollama)
Incremental sync	Re-index everything	Surgical per-file update via GitPython diff
Token savings	No measurement	Built-in dashboard shows exact reduction per query
Graph visibility	Black box	Interactive HTML visualization of every node and edge

---

Prerequisites

• Python 3.10+
• Git — your project must be a git repository (git init if not)
• LLM Provider (choose one):
  • Ollama (local, privacy-first) — install from ollama.com, then pull a model:

    ollama pull llama3
    

  • Claude API (cloud, no local setup) — get API key from console.anthropic.com
  • OpenAI API (cloud, widely available) — get API key from platform.openai.com
  • Claude Code/Cursor (MCP integration) — uses your existing Claude session, zero config!

---

Installation

Option A — Homebrew (Mac/Linux, recommended)

brew tap ramdhavepreetam/nervapack
brew install nervapack

Option B — pipx (any platform, cleanest Python install)

pipx install nervapack

Option C — pip

pip install nervapack

With optional features:

pip install "nervapack[metrics]"    # exact token counting (tiktoken)
pip install "nervapack[dashboard]"  # web dashboard (streamlit + plotly)
pip install "nervapack[mcp]"        # MCP server for Claude Code/Cursor
pip install "nervapack[claude]"     # Claude API support (cloud LLM)
pip install "nervapack[openai]"     # OpenAI API support (cloud LLM)
pip install "nervapack[cloud-llm]"  # Both Claude + OpenAI
pip install "nervapack[all]"        # All features

On first run, chromadb downloads onnxruntime embedding models to your cache and tree-sitter compiles its language bindings. This is a one-time setup (~1–2 min).

---

LLM Provider Options

NervaPack supports multiple LLM providers for document-to-code binding. Choose based on your priorities:

Option 1: Ollama (Default — Privacy-First, Free)

Best for: Privacy-conscious users, offline work, no recurring costs

# One-time setup
brew install ollama
ollama pull llama3
ollama serve

# Use NervaPack (auto-detects Ollama)
nervapack ingest .

Pros: ✅ Free, ✅ 100% private, ✅ Works offline Cons: ❌ ~4GB download, ❌ Uses local resources

---

Option 2: Claude API (Cloud — Fast, High Quality)

Best for: Users who want best results without local setup

# Install Claude support
pip install "nervapack[claude]"

# Get API key from https://console.anthropic.com
export ANTHROPIC_API_KEY=sk-ant-...

# Use NervaPack with Claude
nervapack ingest . --llm claude

Cost estimate: ~$0.25 per 1000 binding calls (Haiku model) Pros: ✅ No local install, ✅ High quality, ✅ Fast Cons: ❌ Sends code to cloud, ❌ Costs money

Available models:

• claude-3-haiku-20240307 (fastest, cheapest — default)
• claude-3-5-sonnet-20241022 (best quality, higher cost)

---

Option 3: OpenAI API (Cloud — Widely Available)

Best for: Users with existing OpenAI accounts

# Install OpenAI support
pip install "nervapack[openai]"

# Get API key from https://platform.openai.com/api-keys
export OPENAI_API_KEY=sk-...

# Use NervaPack with OpenAI
nervapack ingest . --llm openai

Cost estimate: ~$0.15 per 1000 binding calls (GPT-4o-mini) Pros: ✅ Widely available, ✅ Good quality Cons: ❌ Sends code to cloud, ❌ Costs money

Available models:

• gpt-4o-mini (cheapest, good quality — default)
• gpt-4o (best quality, higher cost)

---

Option 4: MCP Delegation (Claude Code/Cursor)

Best for: Users already in Claude Code or Cursor

# Zero setup needed!
# If using NervaPack through Claude Code:
nervapack ingest .
# → Auto-detects MCP context
# → Uses your existing Claude session
# → No separate API key needed

Pros: ✅ Zero config, ✅ Uses existing auth, ✅ No extra cost Cons: ⚠️ Only works in MCP-compatible tools

---

Privacy & Cost Comparison

Provider	Privacy	Setup	Cost (1000 calls)	Use Case
Ollama	🔒 100% Local	Medium	Free	Privacy-first
Claude API	☁️ Cloud	Easy	~$0.25	Quality + convenience
OpenAI API	☁️ Cloud	Easy	~$0.15	Existing OpenAI users
MCP Delegation	☁️ Cloud*	Zero	Included	Claude Code users

*Uses Claude through your existing subscription

---

Quick Start

cd your-project/

# 1. Build the knowledge graph (run once)
nervapack ingest .

# 2. Query for context — see focused results + token savings dashboard
nervapack query "How does authentication work?"

# 3. Visualize the graph with search and community detection
nervapack visualize --enhanced --communities

# 4. Launch interactive web dashboard
nervapack serve

# 5. After modifying files, sync the graph incrementally
nervapack sync .

# 6. Check detailed graph health and analytics
nervapack status --detailed

---

Command Reference

nervapack ingest [PATH]

Scans PATH (default: .) and builds the full knowledge graph.

What happens:

1. tree-sitter parses source files into Classes, Functions, and Imports — exact AST nodes, not text chunks.
2. All .md files are chunked by header hierarchy.
3. Each Markdown chunk is sent to your local Ollama model. If the model identifies a code entity the prose explains, a hard EXPLAINS edge is written into the graph.
4. All nodes are embedded and stored in a local ChromaDB instance (.nervapack/chroma_db).

The initial LLM binding pass is the slowest step. On a large repo with many docs, budget several minutes.

Supported languages (bundled): Python, JavaScript, JSX, TypeScript, TSX

Additional languages (optional extras):

pip install "nervapack[go]"           # Go
pip install "nervapack[rust]"         # Rust
pip install "nervapack[java]"         # Java
pip install "nervapack[c]"            # C / C headers
pip install "nervapack[cpp]"          # C++
pip install "nervapack[ruby]"         # Ruby
pip install "nervapack[csharp]"       # C#
pip install "nervapack[all-languages]" # everything above

---

nervapack query PROMPT

Retrieves context from the graph for a natural-language prompt, then prints a token savings dashboard comparing NervaPack against naive RAG.

What happens:

1. The prompt is embedded and ChromaDB returns the top-3 most semantically similar nodes.
2. Those nodes seed a K-Hop Breadth-First Search (max_hops=1) through the NetworkX graph.
3. Adjacent nodes — including any Markdown docs linked via EXPLAINS edges — are collected into a compressed Markdown snippet.
4. The token efficiency panel is printed showing how many tokens were saved vs. sending the raw files.

Example output:

Running query: How does the CLI work?
Found 3 seed nodes. Traversing graph...

--- Retrieved Context ---
# NervaPack Context Retrieval
## File: src/nervapack/cli.py
### FUNCTION: query (L200-L242)
...
--- End Context ---

╭──────────────  NervaPack Token Efficiency  ──────────────╮
│  Strategy              Tokens   Visual            Relative │
│  Naive RAG (3 files)   12,840   ████████████████  100%    │
│  NervaPack              1,180   █░░░░░░░░░░░░░░░    9.2%  │
│ ──────────────────────────────────────────────────────────│
│  Tokens saved: 11,660   Reduction: 90.8%                  │
│  Cost saved (GPT-4o  $2.50/1M): $0.0292 per query         │
│  Cost saved (Claude Sonnet $3/1M): $0.0350 per query      │
╰───────────────────────────────────────────────────────────╯

"Naive RAG" is defined as the full content of every source file that contains a matched node — the maximum a standard "find relevant files, dump them whole" approach would send to an LLM. The comparison is honest and conservative.

Install nervapack[metrics] for exact token counts via tiktoken. Without it, a character-based estimate is used and marked with ~.

The context output is designed to be pasted directly into an LLM prompt.

📊 Verified Performance: The 90% reduction claim is verified through real-world testing. See BENCHMARKS.md for detailed test results (91.2% average across 5 queries on NervaPack's own codebase). Performance varies based on code quality: 90-99% on clean code, 50-75% on legacy/messy code. Learn more →

---

nervapack visualize [--enhanced] [--communities]

Renders the knowledge graph as an interactive HTML file and opens it in your browser.

Basic usage:

nervapack visualize                          # saves to .nervapack/graph.html
nervapack visualize --output ~/my-graph.html # custom output path
nervapack visualize --no-browser             # generate without opening

Enhanced mode (recommended):

nervapack visualize --enhanced               # adds real-time search
nervapack visualize --enhanced --communities # + community detection

Enhanced features:

• Real-time search — Filter nodes by typing (instant, client-side)
• Path finder — Click two nodes to find and highlight shortest path
• Community detection — Color-coded modules using Louvain algorithm
• Visual highlighting — Matched nodes enlarged, non-matched dimmed

What the visualization shows:

• Node shapes: diamonds = files, dots = all other entities
• Node colors: blue = file, green = function, amber = class, gray = import, lavender = markdown (or community colors)
• Edge styles: solid = DEFINES, dashed = EXPLAINS
• Hover tooltips: type, name, file, line range, and a code preview
• Interactive: drag, zoom, click — spring-force physics layout

The graph is a static HTML file with no external dependencies — share it, open it offline, or embed it in docs.

---

nervapack explore <target> [--hops N]

Extract and visualize a focused subgraph around a specific file, class, or function. Perfect for understanding a specific part of your codebase without the noise of the full graph.

nervapack explore GraphBuilder --hops 2     # explore 2-hop neighborhood
nervapack explore src/cli.py --hops 1       # file-based exploration
nervapack explore "function:parse"          # partial name matching

How it works:

1. Searches for nodes matching the target (by file path, name, or node ID)
2. Performs multi-source BFS to extract N-hop neighborhood
3. Generates enhanced visualization with search enabled

Use cases:

• Understanding class relationships
• Finding related code quickly
• Impact analysis before refactoring
• Onboarding to specific modules

---

nervapack dependencies [file]

Analyze file-level import dependencies, detect circular dependencies, and visualize the dependency graph.

Overall analysis:

nervapack dependencies

Shows:

• Total files and dependency edges
• Circular dependency detection with cycle visualization
• Most depended-on files (top 10)
• Files with most dependencies (top 10)
• Orphan files (no dependencies)
• Interactive hierarchical visualization

Specific file:

nervapack dependencies src/graph/builder.py

Shows:

• Files this file imports
• Files that import this file

Visualization features:

• Color coding:
  • 🔴 Red: Part of circular dependency
  • 🔵 Cyan: Heavily depended on (>5 dependents)
  • 🟡 Yellow: Many dependencies (>5 imports)
  • 🟢 Green: Normal file
  • ⚫ Gray: Orphan (isolated)
• Hierarchical layout — Topological sort when DAG
• Search box — Filter files by name
• Size scaling — Proportional to total degree

Example output:

╭──────────  Dependency Metrics  ──────────╮
│ Total Files              │  127          │
│ Total Dependencies       │  456          │
│ Max Dependency Depth     │  8            │
│ Orphan Files             │  3            │
╰──────────────────────────────────────────╯

⚠ Circular Dependencies Detected: 2 cycle(s)

Cycle 1:
  auth.py
  → user.py
  → session.py
  → auth.py (back to start)

---

nervapack serve [--port N]

Launch an interactive web dashboard with real-time analytics, visualizations, and graph exploration. Requires nervapack[dashboard].

nervapack serve                  # launches on http://localhost:8501
nervapack serve --port 8080      # custom port
nervapack serve --no-browser     # don't auto-open browser

Dashboard features:

• Overview tab — Health score, language distribution, top files
• Analytics tab — Node/edge type breakdown, degree distribution
• Query History tab — Trends, cost savings, keyword frequency
• Graph Explorer tab — Real-time search, statistics

Performance:

• Initial load: <2 seconds
• Subsequent loads: <200ms (cached)
• Real-time filtering and charts

---

nervapack sync [PATH]

Incrementally updates the graph for files changed since the last ingest.

What happens:

1. GitPython diffs your working tree to find modified and deleted files.
2. For each changed file, old graph nodes and ChromaDB vectors are pruned.
3. Only the changed files are re-parsed and re-ingested.

A full ingest on a large codebase can take minutes. sync turns that into a 2–5 second surgical update.

---

nervapack status [--detailed]

Prints the current state of the graph: node count, edge count, and any files that are out of sync with the graph.

Enhanced with --detailed flag:

nervapack status --detailed

Shows comprehensive analytics:

• Health score (0-100) based on documentation coverage, connectivity, and graph density
• Language distribution with visual bars
• Most connected files (top 10 by degree)
• Documentation coverage percentage
• Git sync status with warnings for unsynced files

Example output:

╭──────────── NervaPack Status ────────────────╮
│ Graph Health Score: 85/100 ●●●●●●●●○○        │
│                                              │
│ 📊 Overview                                  │
│   Nodes:      1,247                          │
│   Edges:      3,821                          │
│   Files:        156                          │
│   Functions:    892                          │
│                                              │
│ 📚 Language Distribution                     │
│   Python      ████████████░░░░  62.5%  (98)  │
│   TypeScript  ██████░░░░░░░░░░  35.2%  (55)  │
│                                              │
│ 📖 Documentation Coverage                    │
│   ████████████░░░░░  67.8% (845/1247)        │
╰──────────────────────────────────────────────╯

---

nervapack history [--stats] [--limit N] [--clear]

View query history and analytics. All queries are automatically saved to .nervapack/query_history.jsonl.

nervapack history              # show last 10 queries
nervapack history --limit 20   # show last 20
nervapack history --stats      # aggregate statistics
nervapack history --clear      # clear all history

Statistics include:

• Total queries run
• Average token savings percentage
• Total tokens saved across all queries
• Cost savings (GPT-4o and Claude Sonnet pricing)
• Most frequently queried topics (word frequency analysis)

---

Configuration

NervaPack reads the Ollama model from the LLMSummarizer class (src/nervapack/llm/summarizer.py). To use a different model, set model to any model you have pulled locally:

# src/nervapack/llm/summarizer.py
self.model = "phi3"   # or "mistral", "codellama", etc.

Ollama is expected at http://localhost:11434 (its default). To use a remote Ollama instance, set OLLAMA_HOST:

OLLAMA_HOST=http://my-server:11434 nervapack ingest .

---

Architecture

nervapack ingest .
       │
       ├─ ASTParser (tree-sitter)          16 extensions, 9 languages
       │    └─ ParsedEntity[]: class, function, import
       │
       ├─ GraphBuilder (NetworkX DiGraph)
       │    ├─ Nodes: file, class, function, import, markdown
       │    └─ Edges: DEFINES, EXPLAINS
       │
       ├─ LLMSummarizer (Ollama)
       │    └─ Draws EXPLAINS edges: markdown → code entity
       │
       └─ VectorStore (ChromaDB)
            └─ Embeds node summaries for semantic search

nervapack query "..."
       │
       ├─ VectorStore.search() → seed node IDs
       ├─ GraphRetriever.retrieve_context() → BFS subgraph → Markdown
       └─ TokenMeter → savings vs. naive RAG (tokens, %, cost)

nervapack visualize
       │
       └─ Visualizer (pyvis) → .nervapack/graph.html

Storage layout (inside your project root):

.nervapack/
├── graph.graphml       # NetworkX graph (deterministic structure)
├── graph.html          # Interactive visualization (generated by visualize)
└── chroma_db/          # ChromaDB (semantic embeddings)

Source modules:

Module	Responsibility
nervapack.parser.language_registry	Declarative registry of 16 file extensions and their tree-sitter grammars
nervapack.parser.ast_parser	Tree-sitter parsing → ParsedEntity objects
nervapack.parser.md_chunker	Markdown → header-delimited chunks
nervapack.graph.builder	Build and persist the NetworkX DiGraph
nervapack.graph.vector_store	ChromaDB ingest and semantic search
nervapack.graph.retrieval	K-Hop BFS context extraction
nervapack.graph.visualizer	pyvis interactive HTML export
nervapack.graph.visualizer_v2	Enhanced visualizer with search, communities, path finding
nervapack.graph.dependency_analyzer	File-level dependency analysis and cycle detection
nervapack.graph.analytics	Health scoring, language distribution, coverage metrics
nervapack.graph.query_history	Query tracking and aggregate analytics
nervapack.graph.token_meter	Token counting and savings panel
nervapack.dashboard.app	Streamlit web dashboard with interactive charts
nervapack.llm.summarizer	Local Ollama interface for LLM binding
nervapack.git.tracker	GitPython diff for incremental sync
nervapack.mcp_server	MCP server for Claude Code/Cursor integration

---

Privacy

NervaPack is 100% offline. No code, documentation, or query ever leaves your machine:

• Embeddings are generated by ChromaDB's built-in local model.
• LLM calls go exclusively to localhost:11434 (your Ollama instance).
• All graph and vector data is stored in .nervapack/ inside your project.

Add .nervapack/ to your .gitignore to keep it out of version control.

---

Using NervaPack in LLM Developer Tools

NervaPack ships a built-in MCP server, so any MCP-compatible tool (Claude Code, Cursor, etc.) can use it as a native context provider — no custom code required.

Setup (one-time per project)

1. Install the MCP extra:

pip install "nervapack[mcp]"

2. Build the graph:

nervapack ingest .

3. Add .mcp.json to your project root:

{
  "mcpServers": {
    "nervapack": {
      "command": "nervapack-mcp",
      "description": "NervaPack knowledge graph — query_codebase, graph_status, list_entities"
    }
  }
}

That's it. Reload your MCP-compatible tool and NervaPack's tools appear automatically.

Tools exposed

Tool	What it does
query_codebase(prompt, max_hops?)	Vector search → K-Hop BFS → focused Markdown context + token savings summary
graph_status()	Node/edge counts by type, language breakdown, unsynced file warnings
list_entities(entity_type?, file_path?)	Browse all indexed classes, functions, imports, markdown docs

How Claude Code uses it

Once .mcp.json is in place, Claude Code automatically calls query_codebase before answering questions about the codebase. Instead of reading whole files, it gets a surgical subgraph of only the relevant code — the same token savings you see in the CLI dashboard, applied to every single response.

You:     "How does the sync command decide which files to re-ingest?"
Claude:  → calls query_codebase("sync command file re-ingest logic")
         → gets 1,180 tokens of focused context (vs 12,840 tokens naive)
         → answers precisely, citing exact line numbers

Keeping the graph fresh

# After modifying files
nervapack sync .

# Check if Claude's context is stale
# (graph_status tool reports this automatically)

Python SDK (for building your own tool)

from nervapack.graph.builder import GraphBuilder
from nervapack.graph.vector_store import VectorStore
from nervapack.graph.retrieval import GraphRetriever

graph = GraphBuilder().load_graph()
retriever = GraphRetriever(graph)
results = VectorStore().search("your query", n_results=3)
start_nodes = results["ids"][0]
subgraph = retriever.retrieve_context(start_nodes, max_hops=1)
context = retriever.format_as_markdown(subgraph)
# Inject context into your LLM system prompt

---

nervapack.memory — Conversation Context Extender

Stop re-pasting project context into every new chat.

Every developer using an AI assistant on an ongoing project hits the same wall: Monday you explain the architecture and decisions. Wednesday, new tab, blank slate — you explain them again. nervapack.memory fixes this. It accumulates structured context across sessions and delivers a complete project briefing — 30 days of decisions and conventions — in under 200 tokens.

Session 1:  "Chose JWT for auth — stateless scaling"                → 8 tokens stored
Session 2:  "auth_service issues 15-min tokens"                     → 7 tokens stored
Session 3:  "Deploy: GitHub Actions → staging → manual prod"        → 9 tokens stored
            ...4 weeks later...

memory_recall("project context", budget_tokens=400)
→ 6 items · 171/400 tokens    ← full project briefing, guaranteed ≤ 400 tokens

The bi-temporal model keeps only current truth: old decisions superseded by new ones don't surface. Budget enforcement is a hard invariant — recall never bloats your context window.

Quickstart (5 minutes)

1. Install

pip install "nervapack[memory]"
nervapack-memory init
# ✓ Memory store initialised at .nervapack/memory.db

2. Add to .mcp.json

{
  "mcpServers": {
    "nervapack": {
      "command": "nervapack-mcp"
    },
    "nervapack-memory": {
      "command": "nervapack-memory-mcp"
    }
  }
}

3. Recommended session protocol (add to CLAUDE.md or system prompt)

At the start of every session:
1. Call memory_start_session("task name").
2. Call memory_recall("project context", budget_tokens=400) — this is your briefing.
3. Call memory_recall on any specific topic before working on it.

During the session, store decisions, facts, conventions, and outcomes.
At session end, call memory_end_session("summary").

4. Store a decision with full context

memory_store(
    "Chose JWT over session cookies for auth_service — stateless horizontal scaling",
    kind="decision",
    entities=["auth_service"],
    confidence=0.9,
    rationale="Stateless tokens enable horizontal scaling without shared session store.",
    alternatives_rejected=["server-side sessions", "PASETO"],
)

5. Recall in any future session

memory_recall("project context", budget_tokens=400)

Output:

## Memory recall: "project context" (as of 2026-07-05 · 6 items · 171/400 tokens)

### Decisions
- [d_0019f2] 2026-06-05 · conf 0.90 — Chose JWT for auth_service — stateless horizontal scaling
- [d_0019f4] 2026-06-12 · conf 0.95 — CockroachDB replaces Postgres for geo-distributed writes

### Facts
- [f_0019f3] 2026-06-05 · conf 1.00 — auth_service issues 15-min access tokens with rotating refresh
- [f_0019f7] 2026-06-19 · conf 1.00 — payment_service ~8k txn/day at p99 < 200ms

### Procedures
- [p_0019f5] 2026-06-12 · conf 1.00 — Deploy: GitHub Actions → staging auto → prod manual approval

### Preferences
- [pr_0019f6] 2026-06-12 · conf 1.00 — All new services must expose /health and /metrics endpoints

Token Efficiency: Memory vs. Paste

Approach	Tokens per session	After 20 sessions
Manual paste (architecture doc)	~2,400	~48,000
Manual paste (recent notes)	~800	~16,000
memory_recall	~171	~3,420
Savings	—	93% fewer tokens

Seeding From Existing Notes

Load your decision log or conventions in one pass:

from nervapack.memory import MemoryStore

store = MemoryStore()
decisions = [
    ("Chose JWT for auth_service — stateless scaling", "decision", ["auth_service"]),
    ("CockroachDB replaces Postgres for geo-distributed writes", "decision", ["database"]),
    ("All new services must expose /health and /metrics", "preference", []),
    ("Deploy via GitHub Actions → staging → manual prod approval", "procedure", []),
]
for content, kind, entities in decisions:
    nid = store.add_node(kind=kind, content=content, confidence=1.0)
    print(f"Seeded [{nid}] {content[:50]}")

Data Model

8 node kinds: fact, decision, action, outcome, entity, procedure, preference, session

7 edge kinds: ABOUT, OCCURRED_IN, SUPERSEDES, CONTRADICTS, CAUSED, DERIVED_FROM, TOUCHES

Bi-temporal: every node has valid_from/valid_until (world-time) and recorded_at (learn-time). Supersede closes the old window; rows are never deleted in normal operation.

d_aaa  "session cookies"  valid_from=2026-01-01  valid_until=2026-06-01
  └─[SUPERSEDES]
d_bbb  "JWT"              valid_from=2026-06-01  valid_until=NULL (current)

memory_recall("auth")              → returns d_bbb only (current truth)
memory_recall("auth", as_of=...)  → returns d_aaa only (point-in-time)
memory_timeline("auth")           → returns both (d_aaa marked [superseded])

MCP Tools (17 total)

Tool	Purpose
memory_start_session	Open a named session (e.g. "JWT refactor"); optional namespace param
memory_store	Persist a fact, decision, outcome, procedure, preference, or action; optional namespace param
memory_recall	FTS5 search → graph expansion → scored, budget-capped recall; optional namespace param
memory_about	Entity dossier: all current facts/decisions linked to one entity
memory_why	Explain a decision: rationale, rejected alternatives, caused outcomes
memory_timeline	Chronological trace including superseded versions
memory_end_session	Close session with an outcome summary + queues consolidation
memory_forget	Tombstone (soft) or hard-purge nodes
memory_verify	confirm → confidence +0.1; refute → close + confidence ×0.5
memory_list_sessions	List all sessions with node counts
memory_clear_session	Delete a session and all its nodes
memory_stats	Node counts, DB size, top entities by degree; optional namespace param
memory_for_code	Memories that TOUCH a source file (optionally at a line)
memory_to_code	Code locations a memory node TOUCHES (file, line range, type)
memory_import	Bulk-seed memory from a JSON array of node specs
memory_switch_namespace	Switch active namespace (resets session; all writes go to new NS)
memory_verify_staleness	Scan TOUCHES edges; flag memories where source file changed since stored

Recall Pipeline

query
  ├─ 1. FTS5 BM25 search (exact → prefix* → OR fallback)
  ├─ 2. Graph expansion (0.6× relevance decay per hop, up to 2 hops)
  ├─ 3. Temporal mask (exclude superseded, tombstoned, out-of-window)
  ├─ 4. Confidence filter (min_confidence threshold, default 0.0)
  ├─ 5. Score: relevance × recency × frequency × connectivity
  └─ 6. Budget pack (greedy fill, hard invariant: result ≤ budget_tokens)

CLI

nervapack-memory init                             # create schema
nervapack-memory start-session "Task name"        # open a named session
nervapack-memory stats                            # counts + top entities
nervapack-memory sessions                         # list all sessions
nervapack-memory search "JWT auth"               # FTS search
nervapack-memory show f_0019f2...                 # inspect a node
nervapack-memory forget --node-id f_0019f2...    # tombstone
nervapack-memory forget --entity old_svc --purge # hard-delete
nervapack-memory export --out dump.json           # JSON dump
nervapack-memory import seed.json                 # bulk seed from JSON file
nervapack-memory consolidate                      # deduplicate pending session facts
nervapack-memory consolidate --dry-run            # preview without changes

Storage

Path	When used
.nervapack/memory.db	Default when .nervapack/ dir exists
~/.nervapack/memory.db	Global fallback (created automatically)
$NERVAPACK_MEMORY_DB	Override via environment variable

Combined With the Code Graph

Used together, the two NervaPack servers cover the full context problem:

nervapack-mcp (code graph)          nervapack-memory-mcp (agent memory)
────────────────────────────        ──────────────────────────────────────
"How does auth_service work?"  +    "Why was JWT chosen for auth_service?"
"Which functions call login()?"     "What did we decide last sprint?"
"What does this class import?"      "What's the deploy procedure?"
↓                                   ↓
AST-precise, 91% token reduction    Decision-precise, budget-capped recall

Design

• Standalone SQLite + FTS5 — the code graph (NetworkX + ChromaDB) is immutable between ingest cycles and has no temporal semantics. A separate SQLite file is the right substrate for mutable, session-scoped memory.
• Facts, not chunks — recall returns atomic assertions (8–30 tokens each) with provenance, not transcript segments.
• Bi-temporal, never delete — supersede closes the old valid_until; only memory_forget(purge=True) hard-deletes.
• No network at runtime — all storage, search, and scoring is local. No embeddings at query time.
• Hard token invariant — pack() guarantees the result always fits in budget_tokens. No surprises.

---

Contributing

1. Fork the repo and create a branch.
2. Make your changes with tests where applicable.
3. Open a pull request against master.

Bug reports and feature requests go to the issue tracker.

---

License

MIT — see LICENSE.