Smriti AI live dashboard
External memory for small models.
Inject facts, distract the model, inspect retrieved memories, trace graph facts, verify identity governance, and prove privacy deletion from one presentation-ready cockpit.
Capability map
Every upgrade, visible in one pass
The live UI exercises memory paths directly; the evidence table reads current Gemma 4 CSV artifacts.
Base model
none loaded (fast memory-only dashboard).
Semantic retrieval
Session -> topic -> fact memory with embedding retrieval.
Knowledge graph
Extracts user triples and shows related graph facts.
Identity governance
GIFP drift scoring is computed for each demo response.
Durable backends
JSON, SQLite, Redis, and Postgres are supported.
Encrypted memory
Set SMRITI_MEMORY_KEY to encrypt backend blobs.
Privacy deletion
Delete clears RAM, durable memory, and identity state for a user.
Gemma 4 benchmarks
Dashboard reads current CSV artifacts, not CI smoke results.
HF deployment
Model-style handler is live on Hugging Face.
API + monitoring
FastAPI docs, metrics, Docker, and Grafana assets exist.
AHE harness loop
YAML params, evidence logs, and manifest-backed evolution are auditable.
Production gates
Manifest, privacy/delete, security, holdout, latency, and token gates are visible.
Production-break tests
Malformed input, cross-user isolation, provider outage, prompt injection, and deployment breakers.
Release package
smriti-memory-ai==1.0.9 plus GitHub wheel/sdist artifacts.
No fake claims
Test doubles stay CI-only and never count as production benchmark evidence.
Release and production readiness
Today’s hardening, visible
This panel surfaces the agent hardening layer, production-break suite, file audit, security acceptance, and release packaging state without exposing secrets or private memory.
GitHub release artifacts are versioned; PyPI publish must use a non-exposed token or trusted publishing.
No fake/mock/tiny model, synthetic official benchmark, unsafe harness, secret leak, or cross-user memory path is allowed in production.
Production-break suite
unknown tests, 0 failures, 0 errors.
Production gates
0/0 gates passed for unknown harness.
Security acceptance
OWASP LLM/API risk mapping, auth/RBAC, redaction, delete, and harness security reports.
Operations readiness
0 Grafana dashboards plus runbooks, health checks, and support flow.
Production file audit
0 sensitive findings; 0 review/archive candidates remain non-deleted.
Result provenance
Public claims reject fake, synthetic, or unverifiable benchmark rows.
No fake/mock/tiny model fallback in production.
No official benchmark claim from synthetic or deterministic smoke data.
Production harness promotion requires manifest, privacy, security, latency, token, and holdout gates.
Deletion, auth/RBAC, cross-user isolation, and secret redaction are hard-fail conditions.
Agentic harness evolution
Observable memory knobs
- Config
- configs/harness_params.yaml
- Manifest
- /app/manifests/evolve_manifest.jsonl
- Entries
- 0
- Registered
- 0 harnesses
- Production gate
- review required
- Breaker suite
- missing
| # | Component | Change | Status | Rollback |
|---|---|---|---|---|
No evolution entries yet. Run python benchmarks/collect_evidence.py then python evolve_harness.py. | ||||
Each automated or manual parameter change records before/after config snapshots, a reason, and predicted impact. The base model remains frozen.
One or more promotion safety thresholds failed or evidence is missing.
Industry-standard benchmark check
Internal production gates pass, but these results should not be positioned as full enterprise/industry proof until the warnings below are resolved with larger real-model, real-backend, and user-like benchmark runs.
- Evolved harness ties seed recall; improvement is latency/controls, not recall uplift.
- Validation set is a small deterministic harness benchmark, not an industry-scale external eval.
- Holdout report is missing.
| System | Recall | Precision@K | p50 ms | p95 ms | Token overhead | Delete |
|---|---|---|---|---|---|---|
| No validation table yet. Run quick evaluation. | ||||||
| Model family | Seed recall | Evolved recall | p95 delta | Token delta | Gate |
|---|---|---|---|---|---|
| No cross-model harness report yet. | |||||
Live memory flow
Inject, distract, recall
1 Inject
2 Distract
3 Recall
4 Inspect
5 Delete
Privacy and ops
Backend, deletion, API
- Model mode
- Memory-only / none loaded
- Device
- not loaded
- Training
- not fine-tuned
- Release
- smriti-memory-ai==1.0.9
- PyPI
- release-gated; publish through GitHub secret or trusted publishing
- Guards
- claim provenance review
- Backend
- Json
- Storage
- /app/data/memory
- Encrypted
- ready; set SMRITI_MEMORY_KEY
- Routes
- /chat, /memory/delete, /graph/query, /metrics, /docs
- HF repo
- luciferai-devil/smriti-ai
Deletion clears the in-memory user store, durable backend record, and demo identity state.
Public demo safety: do not enter real PII. Demo memory is automatically deleted after 900 seconds of inactivity.
Gemma 4 benchmark evidence
Saved comparison table
| Config | Recall | Latency | Tokens |
|---|---|---|---|
| Baseline | 0/3 | 4927 ms | 24.8 |
| TF-IDF | 3/3 | 3481 ms | 23.9 |
| Semantic | 3/3 | 2857 ms | 23.8 |
| Semantic + Graph | 3/3 | 2781 ms | 23.9 |
| Semantic + Graph + Identity | 3/3 | 5164 ms | 24.1 |
Historical +31.2% is not claimed as current. Current broader-protocol rerun is shown separately in the README/model card.