Vismaran

विस्मरण (Sanskrit / Hindi: forgetting, oblivion)

Provable right-to-be-forgotten for AI agent memory — graph, vector, and inference log — with a signed deletion receipt a regulator can verify offline.

Why this exists

India's DPDP Rules take full effect 13 May 2027 (no grace period). GDPR Article 17 is the parallel global obligation. Both require, in plain language: a user can demand to be forgotten, you have a clock to do it, the obligation extends to your processors, and you must keep a record proving you did.

A production AI agent's memory of a single person isn't in one place — it's smeared across three layers:

• Graph — structured knowledge-graph memory (entities, relationships, claims). v0 targets Cognee (fronts Neo4j / Kuzu / FalkorDB).
• Vector — embeddings whose source text mentions the subject. v0 targets pgvector.
• Log — the inference + feedback log a self-improvement loop trains on. v0 targets TensorZero (fronts ClickHouse).

Honoring an Article 17 request means erasing across all three, and proving you did. Nobody has built a clean OSS solution. Vismaran is that solution.

Status

v0.1 alpha — active development. The full erase path is in: all three v0 store adapters — Cognee (graph), pgvector (vector), TensorZero (log) — behind an orchestrator that fans out across them in parallel, purges the provenance ledger, and on full success emits a signed Ed25519 deletion receipt (canonical-JSON, tamper-evident, verifiable offline) — fail-loud (no partial-commit receipt) and idempotent. 106 tests pass on every PR — 63 unit + 43 integration against live Postgres + Neo4j + ClickHouse. The vismaran CLI (keygen / erase / verify) that wraps this for operators is next. Legend: ✅ shipped · 🚧 next · 📋 planned.

Capability	v0.1	Roadmap
vismaran_sdk.tag — contextvar subject propagation across async tasks	✅
ProvenanceIndex — Postgres-backed (subject, framework, record_id, ts) ledger	✅
CogneeGraphAdapter — tier-3 Cypher delete via NodeSet subject::<id> (the wedge)	✅
vismaran_sdk.cognee_wrap.add — subject tagging at ingest + provenance recording	✅
Cognee tier-1 (user-scope cognee.forget(user=)) + tier-2 (dataset-scope)		v0.2
PgvectorVectorAdapter — lineage-driven embedding deletion	✅
TensorZeroLogAdapter — 7 ClickHouse tables, ModelInference cascade by inference_id	✅
vismaran_sdk.tensorzero_wrap — vismaran::subject_id tag injection on inference + feedback	📋
Erasure Orchestrator — parallel fan-out, dry-run preview, fail-loud, idempotent	✅
Signed receipt — Ed25519 over a canonical-JSON manifest, tamper-evident, offline-verifiable	✅
vismaran CLI — keygen / erase / verify subcommands	🚧
FastAPI + HTMX demo (Cognee + pgvector + TensorZero)	📋
Crypto-shred mode		v0.2
Anonymize-partial-subject (graph) — re-embed redacted chunks		v0.2
Kuzu / FalkorDB / Neptune graph backends		v0.2
Mem0 / Zep / Graphiti / Letta adapters		v0.3
Qdrant / Weaviate / Milvus vector		v0.3
CopilotKit / Langfuse / OpenTelemetry log adapters		v0.3
TypeScript client		v0.3
Hosted Vismaran		TBD
Backup-erasure		non-v0

Honest about scope: both Cognee and TensorZero have zero open GDPR/RTBF issues today (verified 2026-05-27). Vismaran exists because that gap is real.

Architecture

Five diagrams. Mermaid sources live in docs/architecture/; the PNGs below are checked-in renderings. Full design + DPDP / GDPR Article 17 clause mapping is in SPEC.md.

1. Component overview

What plugs into what: write path (your agent → vismaran_sdk → memory layers + provenance index) and erase path (data subject → Erasure Orchestrator → adapter fan-out → signed receipt → regulator). Framework-agnostic at the protocol level; v0 ships three concrete adapters.

Source: 1-component.mmd

2. Ingest flow

The provenance contract in one sequence diagram. One agent write = one provenance row. Without this, erasure across opaque embeddings is impossible.

Source: 2-ingest.mmd

3. Erasure flow

Happy path (dry-run → confirm → parallel adapter erase → signed receipt → regulator verifies offline) and the failure path (fail-loud, no partial-commit receipt, idempotent retry).

Source: 3-erasure.mmd

4. Threat model

What the receipt provably attests (everything inside Vismaran's trust boundary) vs. what it doesn't (backups, off-platform copies, fine-tunes that already trained on the subject's data, mirrored observability, malicious operator). Anything outside the boundary needs separate controls.

Source: 4-threat.mmd

5. Adapter protocol

The three interfaces (GraphAdapter, VectorAdapter, LogAdapter) and their v0 + roadmap implementations.

Source: 5-adapter.mmd

Quickstart (target — end-to-end flow lands with the demo milestone)

# clone
git clone https://github.com/sambhal-labs/vismaran && cd vismaran

# bring up the stack: neo4j + postgres+pgvector + clickhouse + tensorzero gateway
docker compose up -d

# install + seed Alice/Bob/Carol across all 3 layers
uv sync --extra all --extra demo
make seed

# preview what would be erased
uv run vismaran erase --subject alice@example.com --dry-run

# do it for real → writes receipt.json
uv run vismaran erase --subject alice@example.com

# verify the receipt offline (regulator-grade proof)
uv run vismaran verify receipt.json

Until the orchestrator and receipt signer land, uv run vismaran prints "not implemented yet" for the erase/verify subcommands.

Provenance contract (the only thing your agent has to do)

You can't erase what you didn't trace. Vismaran's SDK gives you 1-line drop-in wrappers:

from vismaran_sdk.tag import with_subject
from vismaran_sdk.cognee_wrap import add as cognee_add
from vismaran_sdk.tensorzero_wrap import inference as tz_inference

async with with_subject("alice@example.com"):
    await cognee_add(text="...")             # tags + records provenance
    await tz_inference(function_name="chat", input=...)  # tags inference with vismaran::subject_id

The tag prefix vismaran:: is namespaced specifically because TensorZero reserves tensorzero:: — see docs/architecture/.

Discussion welcome

• "Why do you need another memory tool?" — Open an issue. Short answer: this isn't memory tooling, it's compliance-grade erasure infrastructure. We don't store your data; we erase from where you do.
• "When is the [Mem0 / Zep / Graphiti / Qdrant / your-framework] adapter landing?" — Track Adapter requests (after the launch).
• "Will Vismaran ever be hosted?" — TBD. v0 is OSS-only. We'll let inbound demand decide.