Continuum

Continuum

Production-grade memory infrastructure for AI agents. Tiered storage (STM / MTM / LTM), first-class supersession and bi-temporal queries, cost-efficient retrieval. Plugs in under whatever reasoner you're using — LangGraph, AutoGen, custom — and provides the memory layer they all lack.

git clone https://github.com/Genkryptos/Continuum.git && cd Continuum
make demo-chat       # 60-second walkthrough, no infrastructure, no API key
make bench-all       # the four benchmarks Continuum is built to win

---

What it is

Most "memory" for LLM agents is one of:

• a vector database (great recall, no notion of "current" vs "stale" facts),
• an append-only chat history (no organization, gets expensive fast),
• a manually-curated profile object (works until the user changes their mind).

Continuum is the operational memory layer: tiered storage so recent conversation, mid-term summaries, and long-term facts are addressable separately; supersession — superseded LTM facts are stamped invalidated_at (never deleted), so every read filters on invalidated_at IS NULL and the system always knows which version is current; and bi-temporal columns — valid_from/valid_to (world-time) plus created_at/invalidated_at (system-time) — so it can answer "what did the user say about X as of date Y?", including retroactive corrections.

It is not a reasoning engine. The framework's value-prop is what the layer below the reasoner should have been doing all along, with benchmark numbers below to back it up.

Why this exists — the headline measurements

The robust, reproducible wins are the deterministic ones (they don't move between runs); the LongMemEval number is honest but carries reader variance.

benchmark	Continuum	naive baseline
Supersession correctness (50 scripted updates)	100 %	38 %
Bi-temporal "as of date Y" (20 scripted timelines)	100 %	75 %
LongMemEval-S (500 Q, judged, gpt-oss-120b)	~74 % (73.6 – 75.6 % across repeated runs)	60.8 % (v1.0) · 34.4 % (May ceiling)
↳ same-setup per-category (control run)	multi-session 62 % · temporal 75 % · KU 64 % · assistant 96 % · user 91 %	—
Retrieval recall @ 4 (200-session synthetic corpus)	55 %	55 % (tied — recency signal absent)
Ingest p50 / session (1 user turn)	0.18 ms + 6 LLM-extraction calls	0.00 ms (raw list)

Sources: LongMemEval-S numbers are documented in the v1 findings report and regenerated with make repro-everything (raw run outputs are gitignored, not committed); synthetic benchmarks from bench/, reproducible via make bench-all (~60 s, no infra, no API key).

Honest headline — read this. On LongMemEval-S the judged number is ~74 % (73.6 – 75.6 % across repeated same-setup runs). gpt-oss-120b has ~±3–5 pp per-category run-to-run variance, so a single run is not a stable estimate: the 76.4 % in earlier v2.0 notes was a favorable draw, and read-side additions (a bounded reflect pass + vote-of-3) are within that noise of the no-additions baseline (73.8 %) — approximately accuracy-neutral overall. Per-category figures swing between runs and should not be read as fixed. The reproducible wins are the two deterministic benches above — supersession and bi-temporal "as of date Y", both 100 % — which don't move. We also disproved four "have the memory layer fix the reader" levers (synthesis, router, distillation, temporal code-math — all net-negative), and learned the methodology lesson that measuring gains only on known failures overstates them. Full write-up: docs/limitations.md · docs/report.md.

v2.0 — retrieval reranking (architecture-native — no new model, no compression). Over-fetch the candidate pool, then a cross-encoder (continuum/retrieval/reranker.py) reorders it and keeps the best 24:

• Lifts recall on the retrieval-bound category — multi-session 57.9% → 69.9% (+12pp, n=133 A/B) — with a +2.3pp temporal bonus; knowledge-update flat; gated off for single-session-preference (it reorders away the rubric turns).
• Honest note: this improves retrieval recall, but on the full-500 judged aggregate it lands within gpt-oss-120b's run-to-run noise of the baseline. The judged number is ~74 % (see the honest headline above), not the 76.4 % an earlier favorable single draw reported. The win here is multi-session retrieval quality, not a lift in the headline accuracy.

v1.1 — 60.8% → 72.8% judged (confirmed full-500, gpt-oss-120b), from two memory-attributable fixes — no new model, no reasoning loop, just surfacing data the schema already had:

• Temporal date-surfacing (the driver): temporal questions were answered with the turn dates stripped from the prompt (the model literally replied "no dates available"). Surfacing each turn's date + the reference "now" lifted temporal 41.4% → 72.9% (+32pp) — confirmed by an n=133 A/B (signal +33pp vs a 0.8pp A/A noise floor).
• Supersession-recency: mark the current (non-invalidated) LTM facts and prefer them over stale turns → knowledge-update ~+4.5pp (n=78 A/B).

Both gated to their categories, so the solved ones (single-session 96–98%) didn't move. Honest note: 90%+ on this benchmark needs a frontier reader — our gains are what the memory layer earns at a fixed mid-tier model. The aggregation-prompt and preference levers were tested and cut as net-neutral / unmeasurable (see findings/roadmap_v1.1.md).

v1.0.0 — Continuum broke the May 2026 ~34% LongMemEval-S ceiling to 60.8% judged — and did it without the iterative reasoning we predicted we'd need. What actually moved the number (a stronger answerer + clean direct retrieval + honest scoring), and the reasoning loop we built and cut as net-negative, are documented in findings/reasoning_loop_2026-06.md. A LOCOMO head-to-head vs Mem0 is preliminary (clean run pending) and not yet a published claim.

The supersession + bi-temporal wins aren't tunable parameters — they're consequences of the schema. Append-only stores and vector databases cannot reach these numbers without re-implementing this architecture on top.

---

The 60-second demo

make demo-chat

Runs a scripted walkthrough where the user:

1. introduces themselves with I live in NYC and just adopted a dog named Rex → Continuum extracts user.location = NYC and user.pets.dog = Rex into LTM.
2. updates: I just moved to Boston → the NYC fact is marked superseded_by the Boston one. Both are still in storage; only Boston is "current".
3. queries /show ltm → current facts only (Boston + Rex).
4. queries /show ltm --all → full history with the supersession edge visible.
5. /query "where do I live" → returns Boston, not NYC, by construction.

The full demo runs in 63 ms with the canned LLM (no API key). It's the canonical show of what bench-supersession proves as a number.

See examples/chat_agent/ for source + the interactive REPL.

---

Architecture in one diagram

┌────────────────────────────────────────────────────────────────────┐
│                       ContinuumSession                             │
│                                                                    │
│   process_turn(user_msg)                                           │
│        │                                                           │
│        ▼                                                           │
│   ┌────────┐    ┌──────────┐   ┌─────────┐   ┌──────────────┐     │
│   │  STM   │ →  │ Retriever│ → │Optimizer│ → │  Responder   │     │
│   │ recent │    │ cosine + │   │ chain   │   │ your LLM     │     │
│   │ turns  │    │ composite│   │ (compress│   │ + ctx        │     │
│   │        │    │ scorer   │   │  budget) │   │              │     │
│   └────┬───┘    └──────────┘   └─────────┘   └──────┬───────┘     │
│        │                                            │              │
│        │ promotion (background queue)               │ reply        │
│        ▼                                            ▼              │
│   ┌────────┐                                  to caller            │
│   │  MTM   │ ← session summaries (LLM-generated)                   │
│   └────┬───┘                                                       │
│        ▼                                                           │
│   ┌────────┐                                                       │
│   │  LTM   │ ← atomic facts + entities + supersession             │
│   │ pgvector│   + bi-temporal columns (valid_from, invalidated_at)│
│   └────────┘                                                       │
└────────────────────────────────────────────────────────────────────┘

Each box is a swappable component. The retriever and optimizer chain are protocol-based; the stores live behind STMProtocol / MTMProtocol / LTMProtocol so the in-memory variant works for local dev and the Postgres variant scales to production.

---

Quick start (no infra)

Use it in Python — the library API (pip install continuum-mcp):

import asyncio
from continuum import Memory

async def main():
    mem = Memory.in_memory()                        # zero-config — no Postgres, no model download
    async with mem:
        await mem.add("I moved to Boston")
        await mem.add("Actually I'm in NYC now")    # supersession handles the update
        print(await mem.recall("where do I live?"))
        print(await mem.current("user", "residence"))  # the resolved, current value
        print(await mem.timeline("residence"))          # bi-temporal history, oldest→newest

asyncio.run(main())

On the Postgres path there is a sixth verb, forget — maintenance for a store that would otherwise only ever grow:

report = await mem.forget(unused_for=timedelta(days=180))   # dry run by default
print(report.matched, report.samples)                        # see it before you mean it
await mem.forget(unused_for=timedelta(days=180), dry_run=False)

It is deliberately timid: only facts that are no longer true (superseded) and untouched for unused_for are eligible, it is scoped to the namespace, capped by limit, and it retires rows the same way supersession does — invalidated_at is set, nothing is deleted, so a policy you regret is recoverable. Nothing prunes on a timer, and it is not exposed as an MCP tool: an agent should not be able to decide to forget things about you.

Use it as an MCP server — plug it into any MCP client (Claude Code, Claude Desktop, Cursor) with zero glue. Memory becomes four tools: remember, recall, current, timeline.

pip install "continuum-mcp[mcp,postgres,embed]"

# point it at a Postgres + pgvector database — migrate once
python -m continuum.db.migrate --dsn "$CONTINUUM_DB_DSN"

# register with Claude Code over stdio (any MCP client works the same way)
claude mcp add continuum -e CONTINUUM_DB_DSN="$CONTINUUM_DB_DSN" -- continuum-mcp

# …or serve over HTTP and connect by URL instead:
#   continuum-mcp --http --port 8000
#   claude mcp add -t http continuum http://127.0.0.1:8000/mcp

Embeddings run locally on bge-m3 — no API key is needed for memory itself. Full transport, backend, and client-config reference: docs/mcp.md.

Supersession + bi-temporal history are strongest on the Postgres path (below); Memory.in_memory() is the zero-setup path for demos and tests. Honest scope and known limits: docs/limitations.md.

From source (demo + benchmarks):

git clone https://github.com/Genkryptos/Continuum.git && cd Continuum

# 1. Install Continuum + its dev tooling
pip install -e .

# 2. See the 60-second story
make demo-chat

# 3. Reproduce the benchmarks
make bench-all

# 4. (Optional) reproduce the v1 LongMemEval-S numbers
#    answerer = gpt-oss-120b via OpenRouter; non-reasoning judge.
#    The dataset is fetched on demand (not committed to the repo).
export OPENROUTER_API_KEY=sk-or-…
make repro-everything

That's the complete loop. No Postgres needed for the demo or any of the benchmarks. The full production path with Postgres+pgvector is documented under continuum/stores/postgres/ and exercised by the integration tests under tests/integration/.

Production path (Postgres) + verifying your .env

cp .env.example .env       # fill in a provider key and/or CONTINUUM_DB_DSN
make db-up                 # start Postgres + pgvector (docker-compose), wait until ready
make db-migrate            # create the pgvector extension + LTM schema
make check-env             # ✓/✗ report: config loads, provider key, DB reachable + schema, smoke
make run                   # interactive chat REPL on a real Postgres-backed session

make run (i.e. python -m continuum.chat) starts an interactive you> REPL on a real ContinuumSession — PostgresSTM + PostgresLTM + the hybrid Retriever + an OpenRouter responder. Every turn persists to Postgres and is indexed for recall, so quitting and re-running with the same --session remembers the conversation. make run keeps the embedder off (LTM sparse/trigram + STM recency, no model download); use make run-full for dense semantic recall (downloads bge-m3, ~2.3 GB, once). Useful flags: --session <id> (named persistent session), --model <id> (OpenRouter model, default openai/gpt-4o-mini), --in-memory (no DB — also make run-mem), --mock (deterministic, no LLM call). Inside the REPL: /help, /search <q>, /stats, /session <id>, /exit. Pass extra flags via make run ARGS="--session alice".

make check-env (i.e. python -m continuum.doctor) is the "did I set this up right?" command — it loads your config, detects which LLM provider key is present (in .env or your shell), probes the configured Postgres DSN for the pgvector extension, and runs one in-memory turn end-to-end. A missing key or unreachable DB is a warning (the offline path still works); a broken config or a session that won't start is a failure (non-zero exit, so it works in CI too). Add --ping to validate each provider key against a live read-only API call, or --full to actually load the embedder.

The compose defaults match .env.example's DSN (postgresql://postgres:postgres@localhost:5432/continuum). If your .env uses different credentials, set POSTGRES_USER / POSTGRES_PASSWORD / POSTGRES_DB (env or shell) before make db-up so the two agree.

---

Make targets

target	what
make demo-chat	scripted 60-second chat-agent walkthrough (in-memory)
make run	interactive chat REPL on a real Postgres-backed session (embedder off)
make run-full	same, with the embedder for dense recall (downloads bge-m3)
make db-up / make db-down	start / stop local Postgres+pgvector (docker-compose)
make db-migrate	apply migrations (pgvector ext + LTM schema)
make check-env	verify your .env: config, provider key, DB+schema, smoke
make bench-ingest	ingest throughput (Continuum vs raw list vs mem0)
make bench-retrieval	retrieval recall@k vs naive cosine
make bench-supersession	the killer feature — 100 % vs 38 %
make bench-bitemporal	"as of date Y" lookups — 100 % vs 75 %
make bench-all	run all four benchmarks in sequence
make bench-locomo	LOCOMO + Mem0 head-to-head (preliminary)
make repro-longmemeval	reproduce the LongMemEval-S evaluation
make repro-everything	reproduce both headline runs (LongMemEval + benches)
make test	full test suite (unit + integration)
make test-fast	unit tests only — no infra required
make check	format + lint + mypy strict
make build	build wheel + sdist into dist/
make build-verify	build, then install + smoke-test in a fresh venv

---

Documentation

Four focused docs under docs/:

• Quickstart — five commands to a working ContinuumSession.
• Architecture — the three tiers, promotion lifecycle, supersession, bi-temporal.
• Config reference — every env-var and YAML knob with defaults.
• Operations — Postgres setup, scaling, observability, common production issues.

If you're new, the recommended path is: this README → make demo-chat → docs/architecture.md → docs/quickstart.md. About 30 minutes if you read everything; 10 if you skim.

Honest evaluation — the LongMemEval-S reports

Two reports, read in order — the second corrects the first, which is the point:

findings/longmemeval_2026-05.md (May) — seven full sweeps across four model families and six retrievers found a hard 32-34% substring ceiling, even at 100% recall, and concluded the bottleneck was multi-hop reasoning: a single-shot retrieve → answer pipeline couldn't exceed it "without an external reasoning loop."

findings/reasoning_loop_2026-06.md (June, v1) — we built that reasoning loop (the IterativeReasoner), A/B'd it, and it was net-negative — so we cut it. v1 broke the ceiling to 60.8% judged with a single-shot pipeline anyway. What actually moved the number:

• a stronger answerer (the dominant lever — we say so plainly),
• the LLM judge revealing substring under-counted paraphrases by +12.4 pp,
• session-aware retrieval + fixing two silent truncation bugs (multi-session ~16% → 55%),
• LTM supersession making knowledge-update work (98.7% recall).

The June report is deliberately honest about what we got wrong (the reasoning-loop prediction), which lever bought which points, where we still lose (temporal-reasoning at 41%), and why the LOCOMO/Mem0 head-to-head is still preliminary rather than a published win. Reproduce both headline runs with make repro-everything.

The throughline that informs Continuum's positioning: a strong model fed clean, complete context beats elaborate scaffolding — be the memory layer that surfaces the right context cheaply; don't try to out-reason the reasoner.

---

Project layout

continuum/                  the framework itself
├── core/                   session orchestration, types, config
├── stores/                 STM / MTM / LTM implementations
│   ├── stm/                in-memory + thread-safe + Postgres
│   ├── in_memory/          in-memory LTM with supersession (eval + local dev)
│   └── postgres/           pgvector-backed MTM + LTM with supersession + bi-temporal
├── retrieval/              composite scorer + BM25 + reciprocal-rank-fusion hybrid
├── extraction/             entity / fact / LLM extractors + cached SmallLLM helper
├── promotion/              Mem0Promoter, triggers, IdleStmFlush
├── optimizer/              token-budget compression chain (5 strategies)
├── scoring/                composite scorer (relevance / importance / recency / confidence)
├── policies/               policy engine + 8 default policies (migration 004)
├── reasoning/              IterativeReasoner — shipped but cut from v1 (tested negative result)
├── embeddings/             embedding service (sentence-transformers)
└── db/                     pgvector upgrade helpers

memory/                     legacy STM engine the framework re-exports (ConversationSTM)

bench/                      memory-operation benchmarks (make bench-all)
├── ingest_throughput.py    Continuum vs raw list vs mem0
├── retrieval_quality.py    recall@k vs naive cosine
├── supersession_correctness.py   100% vs 38%
└── bi_temporal.py          "as of date Y" — 100% vs 75%

evals/                      reproduction harness (datasets fetched on demand)
├── longmemeval/            LongMemEval-S driver (the v1 60.8% result)
└── locomo/                 LOCOMO + Mem0 head-to-head (preliminary)

findings/                   evaluation reports + reproducibility
├── longmemeval_2026-05.md  the May "32% ceiling" report
├── reasoning_loop_2026-06.md   the v1 correction (supersedes it)
└── charts/                 summary-generation scripts

docs/                       quickstart · architecture · config · operations
examples/chat_agent/        the 60-second CLI demo (make demo-chat)
scripts/                    diagnostic-sample + bench-regression + install-verify helpers
migrations/                 numbered Postgres migrations (pgvector, lexical search, policies)

tests/
├── unit/                   per-component unit tests
├── integration/            cross-tier flows + Postgres pgvector
└── acceptance/             phase-completion gates

This release branch ships the essential library + reproduction harness only. Vendored datasets, raw run outputs (results/, bench/results/*.json, .wiki_cache*), and earlier prototype code are intentionally excluded — they're regenerated on demand and gitignored. Full history lives on main.

---

What Continuum is not

• Not a vector database. It uses pgvector under the hood but adds tier semantics, supersession, and bi-temporal queries that pgvector alone doesn't provide.
• Not a reasoning engine. No agentic loop, no multi-hop retrieval, no chain-of-thought orchestration. Use LangGraph, AutoGen, or your own loop on top — Continuum is what they should plug into for memory.
• Not a managed service. It's an open-source Python framework. Self-host on your own Postgres.
• Not a leaderboard-tuned LongMemEval contender. We measured honestly; see the report.

---

Status

Phase 1 (storage + protocols + tests), Phase 2 (extraction + promotion + policies + retrieval), and Phase 3 (benchmarks + demo + findings) are complete. The framework passes strict mypy + full unit / integration / acceptance test gates.

Production readiness (deploy scripts, observability, multi-tenant testing) is the next phase if and when there's interest.

---

License

MIT.

Citation

If Continuum's supersession or bi-temporal benchmarks are useful in your work:

Continuum: production-grade memory infrastructure for AI agents.
https://github.com/Genkryptos/Continuum, 2026.