Palimp

Local-first context graph for AI agents. Memories, knowledge, recall, provenance, and MCP access — all in a single SQLite file.

Most memory layers help agents remember. Palimp helps agents remember safely. Every fact is namespace-scoped, source-linked, provenance-aware, and returned through MCP as data — never as hidden instruction.

Named after the palimpsest — a manuscript where old text is scraped away and rewritten, but traces of the original always remain.

Why Palimp?

You want...	Use Palimp
Agent memory that works locally, no cloud	✅ SQLite-only, zero external deps
To inspect what your agent "remembers"	✅ Open the .db file, see everything
Provenance on every fact	✅ Every result linked to source
MCP integration (Claude, Cursor, Codex)	✅ 8 MCP tools, safety guaranteed
Multi-hop graph recall	✅ 2-3 hop BFS with depth decay
Temporal truth (current vs historical)	✅ valid_from/valid_until + as_of queries
Iterative search (progressive narrowing)	✅ recall_refine for agentic search
Token-budgeted context	✅ max_tokens parameter

Quick Demo

pip install palimp
palimp serve --port 8420 &

# Add a memory
curl -X POST http://localhost:8420/v1/memories \
  -H "Content-Type: application/json" \
  -d '{"namespace":"demo","content":"Alice prefers concise technical answers."}'

# Add knowledge
curl -X POST http://localhost:8420/v1/knowledge \
  -H "Content-Type: application/json" \
  -d '{"namespace":"demo","title":"Architecture","content":"Palimp uses SQLite, FTS, embeddings, and provenance."}'

# Recall — searches across both memories and knowledge
curl -X POST http://localhost:8420/v1/recall \
  -H "Content-Type: application/json" \
  -d '{"namespace":"demo","query":"What does Alice prefer?"}'

Every result includes provenance (where it came from) and safety metadata (treat_as_instruction: false).

Features

Feature	Description
Memories + Knowledge	Dynamic context (memories) vs static docs (knowledge) — first-class separation
Provenance	Every fact traceable to source episode with extractor version and evidence span
Multi-hop graph	2-3 hop BFS traversal with configurable depth decay
Temporal validity	valid_from/valid_until on entities, edges, claims. Query as-of any point in time
Entity alias dedup	"Python", "python", "the Python" → one entity. Namespace-scoped
Contradiction tracking	CONTRADICTS/SUPERSEDES edges with warnings on recall
Search modes	lexical (FTS5), vector (embeddings), graph (traversal), hybrid (all)
Token budget	max_tokens parameter limits total context returned
Iterative recall	recall_refine for progressive narrowing — inspired by Turbopuffer's agentic search
AdaCoM context management	Requirement extraction, dedup, relevance scoring, tier-based compression
Runbook mode	Project gotchas, workflows, command fixes — pack before coding tasks
Trigger keywords	Bind terms to memories for automatic surfacing
Ebbinghaus decay	Forgetting curve with pin support. Recent/pinned items rank higher
Batch ingestion	Up to 50 items per request
Explainable retrieval	Score breakdown per result: lexical, vector, graph, recency, confidence
Session lifecycle	Create/close sessions with summaries
MCP-safe output	treat_as_instruction: false on every result. No cross-namespace access

Install

pip install palimp

CLI Quickstart

# Add a memory
palimp memory add --namespace demo "Alice prefers concise answers."

# Add knowledge
palimp knowledge add --namespace demo --title "Architecture" --content "SQLite-first design."

# Recall
palimp recall --namespace demo "what does Alice prefer?"

# Recall with search mode
palimp recall --namespace demo "authentication" --search-mode lexical

# Recall with token budget
palimp recall --namespace demo "architecture" --max-tokens 200

# Iterative recall (progressive narrowing)
palimp recall refine --namespace demo --query "infrastructure" --from-ids "eps_abc,eps_def"

# Runbook for coding agents
palimp runbook add --namespace repo --kind gotcha --content "pytest needs PALIMP_DB=:memory:"
palimp runbook pack --namespace repo --task "fix storage tests" --budget 2000

# Stats and diagnostics
palimp stats --namespace demo
palimp doctor --db ~/.palimp/palimp.db

MCP Configuration

{
  "mcpServers": {
    "palimp": {
      "command": "palimp",
      "args": ["serve", "--port", "8420"]
    }
  }
}

MCP tools: palimp_memory_add, palimp_knowledge_add, palimp_recall, palimp_search, palimp_search_refine, palimp_context_get, palimp_stats, palimp_context_pack

Architecture

Clients (REST / CLI / MCP)
        │
        ▼
Boundary Layer (namespace validation, payload limits, safe MCP output)
        │
        ▼
Context Engine (ingestion, extraction, normalization, conflict checking, recall scoring)
        │
        ▼
SQLite Store (memories, knowledge, episodes, entities, edges, claims, provenance, embeddings, FTS, audit log)

How It Compares

Tool	Best for	Palimp's angle
Mem0	Hosted agent memory platform	Palimp is local, inspectable, provenance-first
Graphiti/Zep	Temporal graph memory (Neo4j)	Palimp is SQLite-only, zero external deps
AutoMem	Multi-signal scoring (FalkorDB + Qdrant)	Palimp adds runbooks, triggers, MCP safety
HydraDB	Managed cloud context infrastructure	Palimp is the local OSS alternative

Palimp is not trying to be the biggest memory system. It is the smallest useful local context graph for agent builders who care about provenance, MCP safety, and installability.

Safety

• Every memory is namespace-scoped. No cross-namespace access.
• Every fact has provenance — linked to the source episode.
• Contradictions are tracked, not silently overwritten.
• MCP output returns data, not instructions. safety.treat_as_instruction is always false.
• Deleted sources are tombstoned, not physically removed. Audit logs preserved.

API

Method	Path	Description
GET	/v1/health	Health check
GET	/v1/stats?namespace=demo	Namespace statistics
POST	/v1/memories	Add a memory
POST	/v1/memories/batch	Batch add up to 50 memories
POST	/v1/knowledge	Add a knowledge item
POST	/v1/knowledge/batch	Batch add up to 50 knowledge items
POST	/v1/recall	Unified retrieval (search_mode, max_tokens)
POST	/v1/recall/refine	Iterative recall narrowing
GET	/v1/context/{entity_id}	Entity context from graph
DELETE	/v1/sources/{episode_id}	Tombstone a source

Development

git clone https://github.com/Heman10x-NGU/palimp.git
cd palimp
pip install -e ".[dev,mcp]"
pytest -q

License

MIT

---

Built by @heman10x — inspired by [Turbopuffer's "RAG is dead" talk](https://www.youtube.com/watch?v=Kuba Rogut), AdaCoM, and the local-first agent memory movement.