A research prototype exploring augmented personal memory, combining information retrieval, language models, and human-computer interaction.
As large language models increasingly rely on retrieval-augmented generation (RAG) and external memory systems, an important question is how information can be stored and retrieved in a way that feels natural to humans. Conventional information retrieval often assumes that users can formulate queries sufficiently close to the information they are looking for. Human recollection, however, is frequently incomplete and associative: people may remember a person, an event, a concept, or a vague connection without remembering the exact terminology or context.
Membase is a investigation effort exploring how personal memory systems can accommodate this imperfect retrieval process. It combines semantic representations of stored information with lexical and vector search, while accomodating interactive retrieval mechanisms that help users refine their recollections through additional cues and clarification. Rather than attempting to reproduce the biological implementation of human memory, the project explores which of its functional characteristics can inform the design of external memory systems.
The broader motivation is to make information safely offloadable to machines, allowing people to rely less on maintaining every detail in their own memory while retaining the ability to recover it when needed. The design principles investigated in membase may also inform memory architectures for AI agents and more general information retrieval interfaces.
- Information retrieval: combine exact lexical matching with semantic/vector search. Search runs over both the original text and what a model derived from it. The derived layer never quietly replaces the original as the thing being searched.
- Grounding as a retrieval property, not a generation afterthought: the system can refuse to answer. It says so plainly when its own retrieved evidence doesn't actually support a claim.
- LLMs as a provisional inference layer, never a database: every model-derived field is explicitly regenerable. It's layered strictly underneath what the person actually typed, and never overwrites it.
- Human-computer interaction: when is it worth interrupting someone with a clarifying question? When does that turn a two-second note into an interrogation? How much ambiguity should recall resolve silently, versus hand back to the person, versus simply admit it doesn't know?
- Capture of memory units: one submission can describe several unrelated facts. The system splits it into independently retrievable units. The user doesn't do this manually.
- Structurally enforced authority hierarchy: user-typed content outranks what the user later clarified, which outranks what the model inferred, which outranks the fully-derived search index. Correcting or re-deriving a lower layer never touches a higher one.
- Clarification gating: a genuinely ambiguous term can hold a memory out of search until it's answered. This is bounded and skippable, and a real answer becomes part of the memory's stored semantics, not a transient chat message.
- Grounded retrieval with explicit abstention: hybrid search runs over the original text, the derived semantic representation, and chunked attachments. If the evidence doesn't support an answer, the system says so directly ("I don't have a stored memory about this") instead of guessing.
- Iterative recall: a separate, implemented retrieval path that asks one concrete clarifying question when several stored memories could plausibly answer a query, instead of picking one arbitrarily.
- Background "structurer" pass: once a memory is searchable, a background process keeps deepening its semantic representation over time, without blocking anything on it.
- Contradiction surfacing: when two stored memories disagree, that disagreement is flagged for the user rather than resolved automatically or hidden.
flowchart TD
C["Capture<br/>verbatim raw_text, never rewritten"] --> U1[MemoryUnit]
C --> U2[MemoryUnit]
subgraph U1X [" "]
AC["authoritative_content<br/>user-typed"] --> CL["clarifications<br/>user-answered, append-only"]
CL --> SEM["semantic<br/>LLM-derived, regenerable"]
SEM --> IDX["retrieval_text / searchable_text / embedding<br/>fully derived indexes"]
end
U1 --> AC
U1 --> ST{status}
ST -->|ambiguous term/date| AWAIT["awaiting_clarification<br/>NOT searchable"]
AWAIT -->|answered or skipped| PROC[processing]
ST -->|nothing to ask| PROC
PROC --> READY["ready — searchable"]
Authority hierarchy: highest to lowest, what the person typed → what they later clarified → what the model inferred → the fully-derived search indexes built from all of it. Nothing downstream is ever allowed to overwrite something upstream of it. A re-embed or a re-analysis can only touch its own layer, and below.
flowchart TD
W[Write a memory] --> SEG{Segmentation}
SEG --> Q{Ambiguous?}
Q -->|yes| ASK[Ask one clarifying question]
ASK --> ANS[User answers or skips]
ANS --> ENRICH
Q -->|no| ENRICH["Background: enrich + embed"]
ENRICH --> READY[Searchable]
READY -. later .-> QUERY[Ask a question, in different words]
QUERY --> SEARCH["Hybrid search<br/>vector + full-text"]
SEARCH --> EVAL{Evidence enough?}
EVAL -->|no| REFORM[Reformulate] --> SEARCH
EVAL -->|yes| GROUND{Actually grounded?}
GROUND -->|yes| OUT["Answer + citation"]
GROUND -->|no| ABSTAIN["'I don't have a stored<br/>memory about this.'"]
Why abstention is a first-class outcome, not an error case: the retrieval agent is instructed to mark an answer "grounded" only if it actually addresses the question using what was retrieved. If nothing qualifies, the fixed abstention string is returned with no sources. That's the same code path as a real answer, not a fallback or an exception.
docs/ARCHITECTURE.md: the full design
Every abstraction above (Capture, MemoryUnit, the authority
hierarchy, the status gate, the retrieval Agent, RecallSession)
described in full, including the parts not covered here.
docs/DECISIONS.md: the decision log
What was tried, what was reversed, and why, including ideas that evaluation didn't support.
docs/REPORT.md: engineering/evaluation history
Phase-by-phase findings against the real local model, not just unit tests against fakes.
docs/TODO.md: open questions
What this investigation hasn't settled yet.