v2.1.0 — Cognitive Sandwich Goes Local: Claude can now remember before it thinks and verify before it answers

[samvallad33]

Claude should not just remember. It should behave differently because it remembers.

v2.1.0 ships the Cognitive Sandwich: an opt-in Claude Code hook harness that wraps Claude's reasoning path with Vestige memory before the response and local verification after the draft.

Three months ago, Vestige started as a local MCP memory server built around cognitive science instead of brute-force vector search. Since then it has grown into a full cognitive memory layer: Deep Reference reasoning, active forgetting, a live 3D dashboard, Autopilot, and now the local Sanhedrin verifier.

The Problem

Most AI memory systems still behave like retrieval plugins:

1. Store a fact.
2. Retrieve similar facts.
3. Hope the model uses the right one.

That is not enough for long-running Claude Code work. Claude needs memory before it starts reasoning, and it needs a way to catch stale, contradicted, or unsupported claims before the user sees them.

The Solution: The Cognitive Sandwich

v2.1.0 adds two layers around Claude Code:

User prompt
  ↓
TOP BREAD: pre-response hooks
  - Vestige memory graph injection
  - git/CWD state injection
  - fresh dream insight injection
  - lateral-thinking preflight
  ↓
CLAUDE REASONS
  ↓
BOTTOM BREAD: post-response hooks
  - fast veto detector
  - local Sanhedrin verifier
  - synthesis/hedge validator
  ↓
User sees the answer only after the checks pass

The headline piece is the Sanhedrin Executioner. It runs mlx-community/Qwen3.6-35B-A3B-4bit through mlx_lm.server on Apple Silicon, checks Claude's draft against high-trust Vestige evidence, and can veto the answer before it reaches the user.

No Anthropic API call. No cloud round trip. Fully local.

What Changed Since Launch

1. Deep Reference

Vestige now has an 8-stage reasoning pipeline:

hybrid retrieval -> reranking -> spreading activation -> FSRS trust scoring -> temporal supersession -> contradiction analysis -> relation assessment -> reasoning-chain generation.

Instead of returning "10 similar memories," Vestige returns primary evidence, supporting memories, contradicting memories, confidence, trust/staleness notes, and a reasoning scaffold Claude can use.

2. Active Forgetting

Vestige supports explicit top-down suppression.

Not delete. Not demote. Suppress.

Misleading or stale memories can be inhibited without destroying the row. Their retrieval pressure drops, related memories can decay through a Rac1-inspired cascade, and suppression can be reversed within the labile window.

Forgetting is not data loss. It is control over what gets to influence Claude.

3. 3D Memory Dashboard

The dashboard now makes the memory system inspectable: live memory graph, retention states, suppression, contradiction arcs, duplicate concepts, dream insights, activation spread, and reasoning traces.

AI memory should not be a black box.

4. Autopilot

Vestige no longer waits passively for tool calls.

A supervised backend event subscriber routes memory events into the cognitive engine automatically. Memory creation, search, promotion, suppression, and importance scoring can trigger synaptic tagging, predictive memory, activation spread, prospective polling, auto-consolidation, and Rac1 cascade behavior in the background.

A memory system should not just answer queries. It should manage itself.

5. Cognitive Sandwich

The new v2.1.0 hook layer connects all of this to Claude Code's actual response path.

Before Claude thinks, Vestige can inject trusted memory context. After Claude drafts, Vestige can check the draft against high-trust evidence and veto unsupported or contradicted claims.

This is the shift from "Claude has memory" to "Claude is constrained by memory."

Install

Normal MCP server, Apple Silicon example:

curl -L https://github.com/samvallad33/vestige/releases/download/v2.1.0/vestige-mcp-aarch64-apple-darwin.tar.gz | tar -xz
sudo mv vestige-mcp vestige vestige-restore /usr/local/bin/
claude mcp add vestige vestige-mcp -s user

Optional Cognitive Sandwich hook harness:

curl -fsSL https://raw.githubusercontent.com/samvallad33/vestige/v2.1.0/scripts/install-sandwich.sh | sh

The Sandwich is opt-in. Existing MCP users can upgrade without changing their database or tool usage.

Release Health

• Release workflow: green
• CI: green
• Test Suite: green
• 4 release assets uploaded: Apple Silicon macOS, Intel macOS, Linux x86_64, Windows x86_64
• 24 MCP tools
• 30 cognitive modules
• 1,229 Rust release tests verified locally
• 918 dashboard tests verified locally

Thanks

Huge thank you to everyone who opened issues, contributed code, tested installs, challenged the architecture, or starred the repo. Several of the best v2.1.0 changes came directly from community feedback.

Special callout to #36 for pushing on hooks as infrastructure instead of hoping prompt instructions remember to save context. The Cognitive Sandwich is the answer.

Release: https://github.com/samvallad33/vestige/releases/tag/v2.1.0
Docs: https://github.com/samvallad33/vestige/blob/main/docs/COGNITIVE_SANDWICH.md
Repo: https://github.com/samvallad33/vestige

[Bodge-IT]

This will be hard stop for me. Apple only silicon as far as I can tell.

[Jelloeater]

Yeah, @Bodge-IT I agree. It should definitely be optional and also still be able to run on x86. If it cannot, then I'll prob have to migrate away.

[samvallad33]

@Bodge-IT @Jelloeater I’m sorry.

This is fixed in the corrected v2.1.0 hotfix now published at tag v2.1.0 (1d95347). The default installer now:

• activates zero Vestige Claude Code hooks by default
• makes zero automatic model calls by default
• does not wire preflight hooks
• does not wire Sanhedrin / Stop hooks
• does not start MLX or require the 19 GB model / 20+ GB RAM path
• removes old v2.1.0 Vestige hook wiring from ~/.claude/settings.json
• removes the old com.vestige.mlx-server launchd job on reinstall

To clean up an existing v2.1.0 install:

curl -fsSL https://raw.githubusercontent.com/samvallad33/vestige/v2.1.0/scripts/install-sandwich.sh | bash -s -- --force

Opt-in paths are now explicit:

# Optional preflight layer. Includes preflight-swarm.sh, which can call Claude Haiku.
./scripts/install-sandwich.sh --enable-preflight

# Optional Sanhedrin using any OpenAI-compatible endpoint, including x86/Linux/Intel Mac setups.
./scripts/install-sandwich.sh --enable-sanhedrin \
  --sanhedrin-endpoint=http://127.0.0.1:11434/v1/chat/completions \
  --sanhedrin-model=qwen2.5:14b

# Optional Apple Silicon local MLX autostart only for users who explicitly want it.
./scripts/install-sandwich.sh --enable-sanhedrin --with-launchd

Validation completed: full workspace tests pass, installer migration tests cover default cleanup / custom hook preservation / preflight opt-in / Sanhedrin opt-in / combined opt-in, and the local default reinstall now reports no v2.1.0 Vestige hook wiring and no MLX server process.

[Jelloeater]

@samvallad33 Ah great, yeah I just am also using it w/ https://github.com/NousResearch/hermes-agent for its memory storage. Not sure if you have looked at their memory model for inspiration? A lot of folks are using it, vs OpenClaw. https://openrouter.ai/apps/hermes-agent

[samvallad33]

@Jelloeater I did look into Hermes after you mentioned it, and I think you are right that it is worth paying attention to. The main thing I would borrow is the memory-provider lifecycle: initialize, prefetch relevant context before a turn, sync the turn afterward, extract/curate on session end, and keep provider-specific tools additive instead of forcing every user into one storage model. That lines up really well with the pluggable storage work we are tracking in #47: local/serverless sync by default, PostgreSQL/pgvector as an explicit team backend, and no heavy model dependency for normal x86 users.

Where I think Vestige should stay different is the accountability layer: not just storing memories, but showing why something was recalled, whether it is stale or superseded, how it was reinforced/suppressed, and how it synced across machines. Hermes is strong as an agent runtime and provider host; MemPalace/qmd are strong signals that concrete recall/search matters. The path I want for Vestige is to combine those lessons into local cognitive memory that is inspectable, portable, and safe to trust over long-running coding work.

[Jelloeater]

I'm just happy to share 😁

[Bodge-IT]

@samvallad33 Back in the game! thanks Sam and for fb25020 too.