smrti

स्मृति — "that which is remembered"

Non-dual memory for structured knowledge elicitation. A toolkit for completing multi-section applications with AI collaboration, built around four typed memory stores.

Memory Types

Store	File	What it holds
Procedural	tasks.jsonl	Questions, dependencies, status
Episodic	decisions.jsonl	Events, choices, rationale
Semantic	facts.jsonl	Stable facts, preferences
Ephemeral	ephemeral/	Session scratch (not persisted)

JSONL is the source of truth (git-friendly, append-only). SQLite is a regenerable index for fast queries.

Install

pip install advaita-smrti          # core (no dependencies)
pip install advaita-smrti[mcp]     # with MCP server support

Quick Start

cd my-project
smrti init                  # creates .memory/, .mcp.json, .claude/commands/
# Restart Claude Code — 21 memory tools + 6 slash commands are ready

Or use the Python API directly:

from smrti import Memory

mem = Memory(".memory")
task = mem.tasks.create("Answer question", description="...")
mem.tasks.close(task.id)
mem.close()

CLI

smrti init                  # Set up memory in current project
smrti memory status         # Memory layer summary
smrti memory tasks          # List all tasks
smrti memory rebuild        # Repair index from JSONL
smrti memory compact        # Remove old JSONL versions
smrti --version             # Print version

Create Your Own Config

Paste any application into Claude:

Here are questions from my [grant / college app / form].
Create a smrti config JSON with sections, priorities,
and helper text.

[paste questions]

Or use /smrti-config in Claude Code. Validate with python3 validate_config.py.

What Makes This Different

Most memory systems start from conversation. smrti starts from structured forms: sections, questions, dependencies, priorities. This gives it:

• The form itself as procedural memory (no extraction needed)
• Coherence checking across answers
• Dependency-aware ordering
• A clear "done" criterion

Maps to grant applications, clinical intake, legal discovery, insurance claims — anywhere humans complete complex multi-part forms with an AI collaborator.

Documentation

• CONFIG.md — Creating and validating configs
• WORKFLOW.md — CLI, Claude integration, files
• RFCs — Architecture decisions and specs

Inspirations and References

• beads — git-backed task graphs
• quint-code — decision reasoning trails
• ENGRAM (Patel & Patel, 2026) — typed memory stores: episodic, semantic, procedural
• Memory in the Age of AI Agents (Hu et al., 2025) — survey and taxonomy
• Context Engineering: Sessions & Memory (Google, 2025) — whitepaper

See Also

• Zep — temporal knowledge graphs for agent memory
• Letta — filesystem memory that outperforms specialized systems
• Mem0 — structured summarization and conflict resolution
• Pinecone — vector database for semantic retrieval

License

Apache 2.0 — see LICENSE

---

advaita-smrti (अद्वैत-स्मृति) — the tool and the thinker are not separate.