docs(research): add MemPalace vs Hermes memory providers overlap analysis

- Assesses all 8 Hermes external memory plugins (vendor lock-in tiers)
- Documents MemPalace capabilities already in main (sovereign_store + enforcer)
- Identifies overlap gaps: trust scoring, entity resolution, cross-synthesis
- Recommends short-term: enable Holographic provider wired to retrieval enforcer
- Recommends medium-term: Hindsight local as backend for knowledge-graph features
- Clarifies MemPalace as orchestration layer, not just storage

Refs: #419
Closes #419

docs/memory-provider-overlap-analysis-2026-04-29.md

@@ -0,0 +1,204 @@
+# Hermes Memory Providers vs MemPalace — Overlap Analysis & Vendor Lock-in Report
+
+**Issue:** [#419](https://forge.alexanderwhitestone.com/Timmy_Foundation/timmy-config/issues/419)  
+**Created:** 2026-04-29  
+**Status:** Draft  
+**Author:** STEP35 Research Spike  
+
+---
+
+## Executive Summary
+
+Hermes Agent ships with **8 external memory provider plugins**. We already have a custom **MemPalace** system (via `sovereign_store.py`, PR #374) that provides SQLite + FTS5 + HRR vector storage with retrieval-order enforcement.
+
+**Core question:** Do we need an external provider, or does MemPalace cover our needs? Are we creating confusion by running multiple memory systems?
+
+**Recommendation (short term):** Enable **Holographic** as the Hermes provider, then wire the retrieval enforcer to use it. Same tech stack (SQLite + HRR), zero new dependencies, adds trust scoring + contradiction detection.
+
+**Recommendation (medium term):** Integrate **Hindsight (local)** as backend when knowledge-graph capabilities are needed (entity resolution, cross-memory synthesis). MIT license, Docker self-hosted, Ollama-compatible.
+
+**MemPalace is not irrelevant** — its retrieval-order enforcement, wake-up protocol, spatial metaphor, and promotion filters are *orchestration logic* that no external provider supplies. The providers are *storage backends*. MemPalace is the *discipline layer* on top.
+
+---
+
+## Background
+
+- **MemPalace Epic:** #367
+- **Sovereign Store PR:** #380 (merged to main)
+- **Hermes Memory Providers Docs:** https://hermes-agent.nousresearch.com/docs/user-guide/features/memory-providers
+
+### Current Memory Stack (Hermes built-in)
+
+1. `MEMORY.md` — flat file, always active
+2. `USER.md` — user profile, always active
+3. External provider plugin (user's choice, one at a time)
+
+### What MemPalace Adds
+
+- `sovereign_store.py`: SQLite + FTS5 + HRR vectors
+- `retrieval_enforcer.py`: L0→L1→L2→L3→L4 retrieval order
+- `wakeup.py`: Palace-first boot (identity + top rooms in 238 tokens)
+- `promotion.py`: Quality gates before writing to durable memory
+- `scratchpad.py`: Ephemeral session context
+- `identity.txt` + `mempalace.yaml`: Persistent identity + spatial rooms
+- Nightly re-mine cron (Bezalel pattern)
+
+---
+
+## The 8 Hermes Memory Providers — Vendor Lock-in Assessment
+
+### Zero Lock-in (fully self-hosted, open source)
+
+| Provider | License | Storage | Cost | Dependencies | Unique Capability |
+|----------|---------|---------|------|--------------|-------------------|
+| **Holographic** | Ships w/ Hermes (MIT) | Local SQLite | Free | None | HRR algebra, trust scoring, contradiction detection |
+| **Hindsight (local)** | MIT | Embedded PostgreSQL (Docker) | Free | Docker + LLM (Ollama ok) | Knowledge graph, 4-network memory, `reflect` synthesis |
+| **OpenViking** | AGPL-3.0 | Self-hosted server | Free | `openviking` + server | Filesystem hierarchy, tiered retrieval (L0/L1/L2) |
+
+### Partial Lock-in (OSS exists but cloud is default)
+
+| Provider | License | Storage | Cost | Risk |
+|----------|---------|---------|------|------|
+| **Mem0 (self-hosted)** | OSS | Docker (3 containers) | Free | Cloud/OSS feature parity may drift |
+| **Honcho** | OSS option | Cloud or self-hosted | Free/Paid | Self-host path less documented |
+| **ByteRover (local)** | Unclear | Local default | Free | CLI not clearly OSS |
+
+### Full Lock-in (cloud-only, proprietary)
+
+| Provider | Storage | Cost | Risk |
+|----------|---------|------|------|
+| **Supermemory** | Cloud only | Paid | High — no self-host, data on their servers |
+| **RetainDB** | Cloud only | $20/mo | High — proprietary, no exit path |
+
+---
+
+## Overlap Analysis: MemPalace vs External Providers
+
+### What MemPalace Already Does (merged in main)
+
+- ✅ SQLite + FTS5 + HRR vectors (same tech as Holographic)
+- ✅ L0→L1→L2→L3→L4 retrieval order enforcement
+- ✅ Palace-first wake-up (identity + top rooms, 238 tokens)
+- ✅ Quality gates before writing (`promotion.py`)
+- ✅ Ephemeral session scratchpad
+- ✅ Persistent agent identity + spatial room metaphor
+- ✅ Nightly re-mine cron
+
+### What External Providers Add That MemPalace Doesn't
+
+| Capability | Which Provider | MemPalace Gap |
+|------------|----------------|---------------|
+| Knowledge graph with entity resolution | Hindsight | Stores flat facts, no entity linking |
+| Cross-memory synthesis (`reflect`) | Hindsight | No "what do I believe about X across all memories?" |
+| Trust scoring with feedback loops | Holographic | `sovereign_store.py` has no confidence/trust mechanism |
+| Contradiction detection | Holographic | MemPalace can store conflicting facts without noticing |
+| Tiered context loading (100 tok → 2k → full) | OpenViking | MemPalace loads full content, no progressive disclosure |
+| Server-side LLM fact extraction | Mem0, Hindsight | MemPalace extraction is heuristic (regex/length), not LLM-powered |
+| Cross-agent memory federation | Honcho | MP-6 (cross-agent federation) was designed but never implemented |
+
+---
+
+## The Confusion Risk
+
+Hermes already runs **3 memory layers simultaneously**:
+
+1. **Built-in:** `MEMORY.md` + `USER.md` (always active)
+2. **External provider:** whichever plugin is configured (one at a time)
+3. **MemPalace:** our custom `sovereign_store.py` + retrieval enforcer
+
+If we enable an external provider **alongside** MemPalace, agents will have:
+
+- `MEMORY.md` (built-in flat file)
+- `USER.md` (built-in user profile)
+- `sovereign_store.py` (our SQLite palace)
+- External provider tools (e.g., `hindsight_retain`, `hindsight_recall`)
+
+**This is 4 places to store/retrieve memories.** An agent won't know which source of truth to trust.
+
+The retrieval enforcer checks the palace first, but it doesn't know about the external provider. The external provider tools bypass the enforcer entirely.
+
+---
+
+## Decision Options
+
+### Option A: MemPalace Only (status quo)
+
+- Keep `sovereign_store.py` as the sole external memory
+- Accept the gaps (no knowledge graph, no trust scoring, no entity resolution)
+- Zero dependencies, zero lock-in, fully sovereign
+- **Risk:** We miss real capabilities that would make agents smarter
+
+### Option B: Replace MemPalace with Holographic
+
+- Holographic uses the exact same tech (SQLite + FTS5 + HRR) but is more mature
+- Adds trust scoring + contradiction detection we'd have to build ourselves
+- Ships with Hermes — zero extra dependencies
+- **Risk:** We lose the spatial metaphor (rooms/drawers) and retrieval-order enforcement
+
+### Option C: Replace MemPalace with Hindsight (local)
+
+- Most capable option: knowledge graph, entity resolution, reflect synthesis
+- MIT licensed, Docker self-hosted, Ollama for fully offline
+- 91.4% on LongMemEval benchmark (best-in-class for OSS)
+- **Risk:** Adds Docker + LLM dependency for fact extraction. More moving parts.
+
+### Option D: MemPalace as orchestration layer, Hindsight as backend
+
+- Keep retrieval enforcer + wake-up protocol + spatial metaphor
+- Replace `sovereign_store.py` SQLite backend with Hindsight API calls
+- Best of both: our retrieval discipline + their knowledge graph
+- **Risk:** Integration complexity. Two codebases to maintain.
+
+### Option E: MemPalace + Holographic as Hermes provider *(recommended short-term)*
+
+- Enable Holographic as the Hermes external provider (one config line)
+- Keep MemPalace retrieval enforcer but wire it to query Holographic's fact_store
+- Gets us trust scoring + contradiction detection with minimal change
+- Same tech stack (SQLite), zero new dependencies
+- **Risk:** Still two systems, but they speak the same storage language
+
+---
+
+## Recommendation
+
+### Short Term → Option E
+
+Enable Holographic as the Hermes provider and wire the retrieval enforcer to use it.
+
+- Same storage technology (SQLite + FTS5 + HRR) — no migration pain
+- Adds trust scoring and contradiction detection "for free"
+- Zero new dependencies (Holographic ships with Hermes)
+- Minimal integration work
+
+### Medium Term → Option D
+
+When knowledge-graph capabilities become blocking (entity resolution, cross-memory synthesis), migrate the backend to Hindsight local.
+
+- MIT license, fully self-hostable
+- Ollama-compatible LLM for fact extraction
+- Best-in-class LongMemEval performance (91.4%)
+- More moving parts, but tolerable for the capability gain
+
+### Why MemPalace Isn't Obsolete
+
+MemPalace's **orchestration layer** (retrieval enforcer, wake-up protocol, promotion filters, spatial metaphor) is value that no external provider supplies. The providers are *storage backends*. The fix is to **choose a backend that matches our discipline layer**, not discard the discipline.
+
+---
+
+## Action Items
+
+- [ ] Review recommendation with Timmy
+- [ ] Check Holographic integration test suite
+- [ ] Prototype retrieval-enforcer → Holographic fact_store wiring
+- [ ] Evaluate Hindsight local install overhead (Docker + Ollama)
+- [ ] Update `docs/MEMORY_ARCHITECTURE.md` to reflect chosen path
+
+---
+
+## References
+
+- MemPalace epic: #367
+- Sovereign store PR: #380
+- Hermes memory providers: https://hermes-agent.nousresearch.com/docs/user-guide/features/memory-providers
+- Hindsight (MIT, self-hosted): https://github.com/vectorize-io/hindsight
+- OpenViking (AGPL, self-hosted): https://github.com/volcengine/OpenViking