ECHOFORM Ghost Memory

Effectively unlimited long-term memory for any LLM —
zero context tokens · zero weight updates · cryptographic forgetting curve.

Install · How it works · API · Cloud deploy · Architecture · Provenance

If ECHOFORM saves you a context window — drop a ⭐ on the repo.
_{Stars are the only signal that tells us to keep shipping. It takes one click.}

---

What is this?

ECHOFORM is a production memory substrate for AI agents. Existing systems (Mem0, Zep, Letta, …) treat memory as a retrieval problem — fetch chunks, stuff them back into the prompt, watch the bill scale with episode count. ECHOFORM inverts the direction: it stores agent history as a single side-channel FHRR hypervector that pre-biases attention geometry at inference time, without ever consuming a context token or updating a model weight. Every recall comes with a signed forgetting certificate — a closed-form, cryptographically verifiable attestation of what the system retains and what it has provably let go.

That last property is the wedge. It turns the mathematical capacity ceiling of holographic memory from a flaw into the only auditable forgetting primitive in the field — and the only memory layer that can answer GDPR Article 17 with a proof, not a promise.

---

✨ Highlights

• 🧠 Side-channel memory. Agent history is a single 64 KB FHRR vector. /v1/bias returns a residual-stream additive bias in O(1) per call, regardless of how many episodes are stored.
• 📜 Signed forgetting certificate. Every recall ships with an Ed25519 JWS attesting (D, β, episode_count, archive_count, theoretical_recall_curve), canonicalized via RFC 8785 JCS. Verifiable offline with a standalone 100-line verifier.
• 🛡️ GDPR Art. 17 with proof. DELETE /v1/forget purges both the live superposition AND the cold-tier archive (S3 Object Lock GOVERNANCE + STS BypassGovernanceRetention), then signs an ErasureReceipt.
• 🚀 One-click everything. curl … | bash brings up the full prod stack (Postgres + Redis + S3 + worker) and opens a working web UI in your browser. Same stack click-deploys to Render, Fly, DigitalOcean, or any Kubernetes via Helm.
• 🔌 Model-agnostic. Works with Llama-3.1, Mistral, Qwen, and (via the proxy layer) Claude / GPT / Gemini. No weight changes, no fine-tuning.
• 🔍 Observable end-to-end. OpenTelemetry traces + Prometheus RED/USE metrics + structured logs unified by trace_id. Built for Grafana, Loki, Tempo.
• 🎨 Embedded web UI. Vanilla-JS SPA at / — write, recall, view certificates, run GDPR export/erasure, watch live metrics. Zero build step.

---

One-click install (local)

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/OpenAgentic-Labs/echoform-ghost-memory/main/install.sh | bash

Windows (PowerShell):

iwr -useb https://raw.githubusercontent.com/OpenAgentic-Labs/echoform-ghost-memory/main/install.ps1 | iex

The installer verifies Docker, pulls the production image, brings up postgres + redis + minio + api + worker via Compose, waits for /healthz, and opens http://localhost:8080/ in your default browser. A working web UI loads — no code, no curl, no terminal required.

Don't have Docker? Grab it once at docker.com/products/docker-desktop — the installer fails fast with a friendly error if it's missing.

Stop: cd ~/.echoform && docker compose down Upgrade: cd ~/.echoform && docker compose pull && docker compose up -d

---

One-click cloud deploy

Same production image, four managed targets:

Provider	Click-to-deploy	Cost ballpark	Spec file
Render		~$25/mo	deploy/cloud/render.yaml
Fly.io	fly launch --copy-config	~$5/mo	deploy/cloud/fly.toml
DigitalOcean	doctl apps create --spec deploy/cloud/digitalocean-app.yaml	~$24/mo	deploy/cloud/digitalocean-app.yaml
Kubernetes	helm install echoform deploy/helm/echoform	infra-dependent	deploy/helm/echoform/

Every blueprint provisions the same architecture: API + consolidation worker + managed Postgres + managed Redis, with HPA, PDB, ServiceMonitor, and a working /healthz probe. See deploy/cloud/README.md for the production-hardening checklist (KMS for signing keys, S3 archive with GOVERNANCE Object Lock, WAF, key rotation).

---

How it works

┌──────────────────────────────────────────────────────────────────────────────┐
│                                                                              │
│   ┌──────────┐    write       ┌─────────────┐    enqueue                     │
│   │  client  ├───────────────►│  echoform-  ├─────► Redis Streams            │
│   │  (LLM    │                │     api     │                                │
│   │   agent) │   bias         │             │      pull          ┌─────────┐ │
│   │          │◄───────────────┤  (stateless,│                    │echoform-│ │
│   └──────────┘   (no tokens)  │   N×3..10)  │                    │ worker  │ │
│                               └──┬──────┬───┘                    │(single- │ │
│                                  │      │                        │ writer) │ │
│                          read    │      │ write                  └────┬────┘ │
│                          super-  │      │ episode                     │      │
│                          position│      │                             │      │
│                                  ▼      ▼                             │      │
│                          ┌─────────────────────┐    update FHRR       │      │
│                          │     Postgres 16     │◄─────────────────────┘      │
│                          │  · episodic table   │                             │
│                          │  · semantic (FHRR)  │                             │
│                          │  · audit_log        │                             │
│                          └──────────┬──────────┘                             │
│                                     │ spill > capacity                       │
│                                     ▼                                        │
│                          ┌─────────────────────┐                             │
│                          │    S3 / MinIO       │                             │
│                          │  (Ghost Archive,    │                             │
│                          │  Object Lock        │                             │
│                          │  GOVERNANCE)        │                             │
│                          └─────────────────────┘                             │
│                                                                              │
└──────────────────────────────────────────────────────────────────────────────┘

Three lanes:

1. Hot write — POST /v1/remember redacts PII, scans for MINJA-class prompt-injection, inserts into the weekly-partitioned episodes table, enqueues a ConsolidationJob to Redis Streams. Target p99 ≤ 5 ms.
2. Hot read — POST /v1/bias loads the tenant's 64 KB FHRR superposition (asyncio-locked TTL cache), encodes the query, takes one complex inner product, returns the residual-stream bias + certificate_id. O(1) in episode count. Target p99 ≤ 20 ms.
3. Cold consolidation — A single-writer worker pulls batches from Redis, runs the consolidation pass (Mamba-based in v0.2; IdentityConsolidator stub in v0.1), writes a new superposition row with optimistic CAS, emits a drift metric. Decoupled from the request path; the API never blocks on it.

Full Architecture Decision Record: _artifacts/phase3/architecture-adr.md. Module dependency DAG enforced by .importlinter.

---

The forgetting certificate

ECHOFORM's wedge. The certificate is a detached Ed25519 JWS over a JCS-canonicalized payload:

{
  "deployment_id":          "acme-prod-eu",
  "signing_key_id":         "01J9F3RXK4M7Z2B5W6XE0YQVQR",
  "issued_at":              "2026-05-19T11:23:45Z",
  "D":                      8192,
  "beta_rational":          { "num": 3, "den": 100 },
  "episode_count":          47291,
  "archive_count":          12055,
  "n_star_certified":       909,
  "theoretical_recall":     "erfc(sqrt(n * beta * log(D) / D))",
  "adapter_sha256":         "7f2c…",
  "coprocessor_sha256":     "ae91…",
  "proxy_calibration_eps":  0.012,
  "post_erasure_cosine_max":0.043,
  "enclave_quote_b64":      null
}

Why this matters:

Property	Mem0 / Zep / Letta / Graphiti	ECHOFORM
Forgetting curve published?	❌ (empirical, opaque)	✅ closed-form, derivable from (D, β, n)
Offline verifiable?	❌	✅ Ed25519 JWS, RFC 7515 + RFC 8785
Bound to model + adapter hash?	❌	✅ adapter_sha256 + coprocessor_sha256
Survives alg=none attack?	n/a	✅ rejected at ECHOFORM layer before any JWT call
Covers cold-tier erasure?	❌	✅ S3 Object Lock GOVERNANCE + STS bypass IAM
Pinned crypto stack?	n/a	✅ pyjwt ≥ 2.12 (CVE-2026-32597 patched), PyNaCl, JCS

Standalone verifier (no echoform dependency, runnable by auditors): echoform/security/certificate/verifier.py.

---

The API

Python SDK

import ghost_memory as gm

mem = gm.GhostMemory(endpoint="https://your-deploy.example.com")

# Write
mem.remember(
    agent_id="customer-success-bot",
    content="User asked about Q3 revenue forecast on the discovery call.",
    subject_pseudonym="cust_1f3a",
)

# Recall — returns the residual-stream bias + certificate id
bias = mem.bias(
    agent_id="customer-success-bot",
    query="what did the customer ask about previously?",
)
print(bias.certificate_id, bias.vector_bytes[:16].hex())

# Compliance
cert    = mem.certificate()                # current ForgettingCertificate JWS
export  = mem.export(subject_pseudonym="cust_1f3a")        # GDPR Art. 15
receipt = mem.forget(                                       # GDPR Art. 17
    subject_pseudonym="cust_1f3a",
    audit_reason="DSAR-2026-0042",
)

Raw HTTP

# Write
curl -X POST https://localhost:8080/v1/remember \
  -H 'Content-Type: application/json' \
  -H 'Idempotency-Key: 01J9F3R…' \
  -d '{"agent_id":"demo","content":"hello world","ts":1747641600}'

# Bias
curl -X POST https://localhost:8080/v1/bias \
  -H 'Content-Type: application/json' \
  -d '{"agent_id":"demo","query":"what did I say?"}'

# Certificate
curl https://localhost:8080/v1/certificate

# Erase
curl -X DELETE https://localhost:8080/v1/forget \
  -H 'Content-Type: application/json' \
  -d '{"subject_pseudonym":"cust_1f3a","audit_reason":"DSAR-2026-0042"}'

Full OpenAPI 3.1 spec at /openapi.json once the server is running. Interactive Swagger at /docs. ReDoc at /redoc.

---

What's in the box (v0.1)

Module	What	Status
echoform.core.fhrr	FHRR bind/unbind/bundle/capacity at D=8192 complex64	✅ real
echoform.core.adapter	Llama-3.1-70B residual-stream bias (HF forward hook)	✅ real
echoform.storage.episodic	Postgres 16 + pgcrypto + pg_partman 5.2	✅ real
echoform.storage.semantic	Optimistic CAS + asyncio-locked TTL cache + safetensors	✅ real
echoform.api.public	FastAPI 0.136.1 — 9 stable endpoints + web UI mount	✅ real
echoform.sdk	Async + sync httpx clients + ghost_memory façade	✅ real
echoform.security.certificate	Ed25519 + JCS canonicalization + key rotation	✅ real
echoform.safety	PII redactor + 55 MINJA payloads + poison_scan_gate	✅ real
echoform.observability	structlog + OpenTelemetry + Prometheus	✅ real
echoform.core.compression	Bag-of-tokens FHRR (Mamba-2 → v0.2)	🟡 stub
echoform.core.consolidation	IdentityConsolidator (HiCL → v0.2)	🟡 stub
echoform.storage.archive	LocalArchive + S3 stub with COMPLIANCE refusal	🟡 v0.2 S3 wire-up
echoform.api.proxy	Returns HTTP 501 (Anthropic + OpenAI + vLLM in v0.2)	🟡 stub

See docs/roadmap.md for the v0.2 / v1.0 line.

---

Project structure

echoform-ghost-memory/
├── README.md  LICENSE  NOTICE  CHANGELOG.md  SECURITY.md  GOVERNANCE.md
├── BUILT_FROM.md             ← provenance trail
├── pyproject.toml  tasks.py  Dockerfile  docker-compose.yml
├── install.sh  install.ps1   ← one-click local installer
├── echoform/                 ← the package (82 .py files)
│   ├── types.py              ← canonical types (leaf import)
│   ├── core/{fhrr, compression, adapter, consolidation}
│   ├── storage/{episodic, semantic, archive}
│   ├── api/{public, proxy}
│   ├── sdk/, cli.py
│   ├── observability/, security/certificate/, safety/
│   └── config/settings.py
├── ghost_memory/             ← 5-line public façade
├── tests/                    ← 46 tests across 7 categories
│   ├── unit/  property/  integration/  smoke/  perf/  security/  fuzz/
│   └── conftest.py
├── deploy/
│   ├── k8s/{base, overlays/{dev,prod}, iam}      ← Kustomize
│   ├── cloud/{render.yaml, fly.toml, …}          ← managed-platform blueprints
│   └── helm/echoform/                             ← Helm chart with subcharts
├── docs/
│   ├── architecture.md  ops-runbook.md  security-baseline.md
│   ├── observability.md  data-classification.md
│   ├── glossary.md  governance.md  roadmap.md  known-issues.md
│   └── closed-model-gap.md  assets/
├── examples/                 ← 4 runnable examples
└── .github/workflows/        ← ci.yml, release.yml (cosign), security.yml

---

Production posture

• Container. Multi-stage Dockerfile, non-root UID 10001, distroless-style runtime, signed via Sigstore Cosign on every release.
• CI matrix. ubuntu × macos × windows × python-3.11, python-3.12. Lint (ruff), typecheck (mypy strict on critical paths), unit + property + integration + smoke + security, dep audit (pip-audit), license guard (pip-licenses --fail-on=GPL;LGPL;AGPL;…), gitleaks on every PR, import-DAG enforcement via import-linter.
• Observability. OpenTelemetry SDK 1.41, OTLP gRPC default, trace_id in every log, tenant_id only on trace exemplars (cardinality budget enforced).
• K8s defaults. HPA 3-10, PDB minAvailable=2 on api / 1 on worker, NetworkPolicy default-deny + explicit egress, all three probes (startup + liveness + readiness), securityContext with runAsNonRoot, readOnlyRootFilesystem, drop ALL caps, seccompProfile: RuntimeDefault.
• Compliance. GDPR Art. 15 export + Art. 17 erasure + Art. 20 portable format. EU AI Act Art. 12 audit log retained 6 months minimum. Data classification table in docs/data-classification.md.

---

Roadmap

v0.2 (next):

• Real Mamba-2 130M consolidation rebind (replaces IdentityConsolidator)
• api.proxy Anthropic + OpenAI + vLLM adapters
• Spectral-entropy MINJA detector (replaces regex tier)
• AWS KMS / HashiCorp Vault signing-key backend
• /v1/audit/{event_id} endpoint
• Locust-driven cross-process p99 perf gate

v1.0:

• TEE attestation (AMD SEV-SNP / AWS Nitro Enclaves) — Sovereign SKU
• pgBackRest continuous WAL streaming + PITR
• CBOR / COSE certificate payloads
• Photonic / neuromorphic substrate support (D ≥ 10⁶)

See docs/roadmap.md for the full sequencing.

---

Contributing

Pull requests welcome. We use the Developer Certificate of Origin — git commit -s on each commit; no CLA. Open an issue first for non-trivial changes; architecture changes need an ADR under docs/adr/ per GOVERNANCE.md.

git clone https://github.com/OpenAgentic-Labs/echoform-ghost-memory
cd echoform-ghost-memory
python -m pip install -e ".[safety]"
python tasks.py smoke                       # ✅ < 30 s
python tasks.py test                        # full suite (needs PG + Redis)

All contributors must follow CODE_OF_CONDUCT.md (Contributor Covenant 2.1). Vulnerability reports go to the channel in SECURITY.md.

---

Citation

If you build on ECHOFORM in academic work, please cite:

@software{echoform_ghost_memory_2026,
  title  = {ECHOFORM Ghost Memory: A side-channel forgetting-certified memory substrate for LLM agents},
  author = {{The ECHOFORM Authors}},
  year   = {2026},
  url    = {https://github.com/OpenAgentic-Labs/echoform-ghost-memory},
  version= {0.1.0},
}

---

License

Apache License 2.0 — see NOTICE for third-party attributions (PyTorch, FastAPI, asyncpg, OpenTelemetry, PyJWT, PyNaCl, structlog, mamba-ssm, …). Compliant deps only — cargo-deny / pip-licenses gates AGPL/GPL/LGPL/SSPL/BUSL in CI.

---

Built from

This codebase was assembled in 8 phases (intake, spec crystallization, paired 30-agent dispatch, architecture council, adversarial critic review with veto floor, assembly, pre-flight integration check, end-to-end test verification) per the mind-build@v2 workflow. The provenance trail — dossier SHA-256, archetype pairings, conflict resolutions, Phase-3.5 critic confidence scores, all 10 must-fix items with their resolution — lives in BUILT_FROM.md.

---

Star history

_{Watching this graph climb is how we know the work matters. If you’ve read this far,
the ⭐ star button takes one click — thank you.}

---

_{Made with ∞ by OpenAgentic Labs.}