Igor Lins e Silva
6509071b8e
Default search behavior is unchanged. Opt-in candidate_strategy="union"
also pulls top-K BM25-only candidates from sqlite FTS5 and merges them
into the rerank pool, catching docs with strong BM25 signal that the
vector index didn't surface in the over-fetch window.
Motivation
----------
The current hybrid path gathers candidates from the vector index only
(n_results * 3 over-fetch), then BM25-reranks within them. When the
query embeds close to the wrong content semantically, the right doc
never enters the rerank pool — *no matter how wide the over-fetch*.
Tested on a ~6K-document mixed corpus (knowledge prose + short structured
records): at *30x* over-fetch (~5% of the corpus) the target doc still
didn't surface for narrative-shaped queries targeting terminology guides.
Wider over-fetch isn't the answer; widening the pool's *source* is.
Concrete failure mode: a narrative-shaped query embeds close to records
sharing the same operational vocabulary (other narrative entries in the
corpus). A terminology / style guide is BM25-strong for the query
(rare keywords the guide repeats) but vector-distant. Vector-only
candidates don't include it; BM25 never gets to rerank it. The hybrid
path produces 0.00 recall on a probe that pure BM25 alone scores 1.00 —
the hybrid is worse than its component on the same input.
Behavior change
---------------
* New parameter ``candidate_strategy: str = "vector"`` on ``search_memories``.
- ``"vector"`` (default): historical behavior, no change.
- ``"union"``: also fetch top ``n_results * 3`` candidates via the
existing ``_bm25_only_via_sqlite`` helper, dedupe by source_file,
merge into the rerank pool. BM25-only candidates carry
``distance=None`` so they're scored on BM25 contribution alone
(vec_sim coerces to 0).
* ``_hybrid_rank`` now handles ``distance=None`` explicitly, scoring
such candidates as vector-unknown (vec_sim=0) rather than treating
it as max-distance via shim.
* New strategies register via ``_CANDIDATE_MERGERS``; dispatch is in
``_apply_candidate_strategy`` so ``search_memories`` stays under the
C901 complexity ceiling.
Bench numbers (~6K-doc internal mixed corpus, recall@10, 5 probes spanning
policy-exception lookup, temporal-decay, style retrieval, set-difference,
and pattern-recognition):
baseline ("vector") "union"
policy-exception probe 0.00 0.50 +0.50
temporal-decay probe 0.17 0.50 +0.33
style-retrieval probe 0.00 1.00 +1.00 (PASSES)
set-difference probe 0.00–0.06 0.06–0.09 ~
pattern-recog probe 0.64 (stable) 0.50–0.71 variance, typ. +0.07
macro recall 0.16–0.17 0.51–0.56 +0.34 to +0.40
The pattern-recog variance points at a related issue worth a separate PR:
``_hybrid_rank`` computes BM25 IDF over the candidate set. Adding new
candidates re-normalizes BM25 for *existing* candidates non-monotonically.
Stable corpus-wide BM25 would remove this. Out of scope here.
Tests
-----
``tests/test_hybrid_candidate_union.py`` (6 tests, all pass):
- default behavior unchanged (explicit ``"vector"`` matches default)
- ``"union"`` surfaces a BM25-strong vector-distant doc
- ``"union"`` doesn't drop docs ``"vector"`` would have found
- empty-palace handling
- invalid ``candidate_strategy`` raises
- ``_hybrid_rank`` tolerates ``distance=None``
Existing ``test_hybrid_search.py`` (5) and ``test_searcher.py`` (27) pass.
Performance note
----------------
Each ``"union"`` query adds one sqlite open + FTS5 MATCH + metadata
fetch (via the existing ``_bm25_only_via_sqlite`` helper, which already
runs as the ``vector_disabled`` fallback path so the code is
well-trodden). Per-query overhead is small but unmeasured at corpus
scale. Default stays ``"vector"`` until a maintainer characterizes the
cost.
Caution
Scam alert. The only official sources for MemPalace are this GitHub repository, the PyPI package, and the docs site at mempalaceofficial.com. Any other domain — including
mempalace.tech— is an impostor and may distribute malware. Details and timeline: docs/HISTORY.md.
MemPalace
Local-first AI memory. Verbatim storage, pluggable backend, 96.6% R@5 raw on LongMemEval — zero API calls.
What it is
MemPalace stores your conversation history as verbatim text and retrieves it with semantic search. It does not summarize, extract, or paraphrase. The index is structured — people and projects become wings, topics become rooms, and original content lives in drawers — so searches can be scoped rather than run against a flat corpus.
The retrieval layer is pluggable. The current default is ChromaDB; the
interface is defined in mempalace/backends/base.py
and alternative backends can be dropped in without touching the rest of
the system.
Nothing leaves your machine unless you opt in.
Architecture, concepts, and mining flows: mempalaceofficial.com/concepts/the-palace.
Install
pip install mempalace
mempalace init ~/projects/myapp
Quickstart
# Mine content into the palace
mempalace mine ~/projects/myapp # project files
mempalace mine ~/.claude/projects/ --mode convos # Claude Code sessions (scope with --wing per project)
# Search
mempalace search "why did we switch to GraphQL"
# Load context for a new session
mempalace wake-up
For Claude Code, Gemini CLI, MCP-compatible tools, and local models, see mempalaceofficial.com/guide/getting-started.
Benchmarks
All numbers below are reproducible from this repository with the commands
in benchmarks/BENCHMARKS.md. Full
per-question result files are committed under benchmarks/results_*.
LongMemEval — retrieval recall (R@5, 500 questions):
| Mode | R@5 | LLM required |
|---|---|---|
| Raw (semantic search, no heuristics, no LLM) | 96.6% | None |
| Hybrid v4, held-out 450q (tuned on 50 dev, not seen during training) | 98.4% | None |
| Hybrid v4 + LLM rerank (full 500) | ≥99% | Any capable model |
The raw 96.6% requires no API key, no cloud, and no LLM at any stage. The hybrid pipeline adds keyword boosting, temporal-proximity boosting, and preference-pattern extraction; the held-out 98.4% is the honest generalisable figure.
The rerank pipeline promotes the best candidate out of the top-20
retrieved sessions using an LLM reader. It works with any reasonably
capable model — we have reproduced it with Claude Haiku, Claude Sonnet,
and minimax-m2.7 via Ollama Cloud (no Anthropic dependency). The gap
between raw and reranked is model-agnostic; we do not headline a "100%"
number because the last 0.6% was reached by inspecting specific wrong
answers, which benchmarks/BENCHMARKS.md flags as teaching to the test.
Other benchmarks (full results in benchmarks/BENCHMARKS.md):
| Benchmark | Metric | Score | Notes |
|---|---|---|---|
| LoCoMo (session, top-10, no rerank) | R@10 | 60.3% | 1,986 questions |
| LoCoMo (hybrid v5, top-10, no rerank) | R@10 | 88.9% | Same set |
| ConvoMem (all categories, 250 items) | Avg recall | 92.9% | 50 per category |
| MemBench (ACL 2025, 8,500 items) | R@5 | 80.3% | All categories |
We deliberately do not include a side-by-side comparison against Mem0, Mastra, Hindsight, Supermemory, or Zep. Those projects publish different metrics on different splits, and placing retrieval recall next to end-to-end QA accuracy is not an honest comparison. See each project's own research page for their published numbers.
Reproducing every result:
git clone https://github.com/MemPalace/mempalace.git
cd mempalace
pip install -e ".[dev]"
# see benchmarks/README.md for dataset download commands
python benchmarks/longmemeval_bench.py /path/to/longmemeval_s_cleaned.json
Knowledge graph
MemPalace includes a temporal entity-relationship graph with validity windows — add, query, invalidate, timeline — backed by local SQLite. Usage and tool reference: mempalaceofficial.com/concepts/knowledge-graph.
MCP server
29 MCP tools cover palace reads/writes, knowledge-graph operations, cross-wing navigation, drawer management, and agent diaries. Installation and the full tool list: mempalaceofficial.com/reference/mcp-tools.
Agents
Each specialist agent gets its own wing and diary in the palace.
Discoverable at runtime via mempalace_list_agents — no bloat in your
system prompt:
mempalaceofficial.com/concepts/agents.
Auto-save hooks
Two Claude Code hooks save periodically and before context compression: mempalaceofficial.com/guide/hooks.
For per-message recall on top of the file-level chunks the hooks produce,
run mempalace sweep <transcript-dir> periodically — it stores one
verbatim drawer per user/assistant message, idempotent and resume-safe.
Requirements
- Python 3.9+
- A vector-store backend (ChromaDB by default)
- ~300 MB disk for the default embedding model
No API key is required for the core benchmark path.
Docs
- Getting started → mempalaceofficial.com/guide/getting-started
- CLI reference → mempalaceofficial.com/reference/cli
- Python API → mempalaceofficial.com/reference/python-api
- Full benchmark methodology → benchmarks/BENCHMARKS.md
- Release notes → CHANGELOG.md
- Corrections and public notices → docs/HISTORY.md
Contributing
PRs welcome. See CONTRIBUTING.md.
License
MIT — see LICENSE.