---
type: analysis
slug: research-everos-traces-memory-into-langfuse-confidence-and-cost-take-side-paths
title: "EverOS traces memory into Langfuse; confidence and cost take side paths"
description: "EverOS 1.2.0 added native OpenTelemetry export for store, extract, search, and reflection, and Langfuse documents how those spans land. Recall confidence still posts to Langfuse's scores API rather than OTLP, and dollar cost is computed by Langfuse from token usage."
published_at: 2026-11-07T23:00:00.000Z
topics: ["memory","search","observability","research"]
source_urls: ["https://github.com/EverMind-AI/EverOS/releases/tag/v1.2.0","https://raw.githubusercontent.com/EverMind-AI/EverOS/main/CHANGELOG.md","https://langfuse.com/integrations/other/everos","https://pypi.org/project/everos/","https://raw.githubusercontent.com/EverMind-AI/EverOS/main/src/everos/config/default.toml","https://github.com/EverMind-AI/EverOS/commit/ab5cf0447ea5eff1338326dab1a12b3052bb77f7","https://pypi.org/project/everos/1.2.0/","https://langfuse.com/integrations/native/opentelemetry","https://raw.githubusercontent.com/EverMind-AI/EverOS/main/examples/langfuse/README.md","https://github.com/EverMind-AI/EverOS/commit/357f619c64c5ce03e4c25b76f19e08f0528a0588","https://raw.githubusercontent.com/EverMind-AI/EverOS/main/src/everos/core/observability/tracing/scores.py","https://github.com/EverMind-AI/EverOS/commit/4e13f7881e5bad0d20b31da62083fb6ff0b40eb0","https://github.com/EverMind-AI/EverOS/releases/tag/v1.2.1"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
The flagged post said EverOS now speaks OpenTelemetry natively, so memory store, recall, confidence, conflict-merge, and cost events appear in Langfuse. A social mention is not a contract, and the newsroom was right to ask for a release, a schema, and an integration path.

Those now exist. EverOS is an Apache-2.0 local-first memory runtime (Markdown plus SQLite and LanceDB) published on PyPI. GitHub tagged 1.2.0 on 24 July 2026 with native OTLP export of memory operations. Langfuse published a matching integration page whose lead paragraph matches the post. Current PyPI is 1.5.1 and still lists the optional otel extra. The remaining builder work is to treat memory as newly observable, then distinguish OTLP spans from Langfuse-only scores and Langfuse-computed cost.

## What we checked

- **Is there a public EverOS release note or documentation for the Langfuse/OTel integration?** confirmed. Yes. EverOS 1.2.0, tagged 24 July 2026, added native OpenTelemetry tracing of memory operations to any OTLP backend, with Langfuse named as the example. Langfuse published a matching integration page, and the EverOS repo documents the same config in examples/langfuse. [1][2][3][9]
- **What exact telemetry schema and Langfuse integration path are available to builders?** confirmed. Install everos[otel], set [observability] with Langfuse keys or a raw OTLP/HTTP endpoint, and traces appear without app-side instrumentation. Writes, extracts, searches, embeddings, OME strategies, and consolidations map to named observations; recall scores attach separately via Langfuse POST /api/public/scores. [3][9][5][11][7]
- **Does EverOS export recall confidence over OpenTelemetry, as the post implies?** refuted. Search spans are OTLP. Recall scores are a Langfuse-specific REST object posted off the request path to /api/public/scores, and they only fire when Langfuse credentials and emit_recall_scores are set. A generic OTLP collector will not receive those scores. [11][5][10]
- **Does EverOS emit the dollar cost of remembering?** partly confirmed. Generation and embedding observations carry model and token usage so Langfuse can price them in its model-usage views. Langfuse states EverOS sends no cost of its own, so a non-Langfuse backend sees usage attributes, not a billed price. [3][8]

## What we found

- EverOS 1.2.0, published 24 July 2026, added native OpenTelemetry tracing of add/flush, memcell boundary, episode extraction, search, and OME reflection to any OTLP backend, with Langfuse named as the example. The CHANGELOG repeats that contract, and it remains present through current 1.5.1. Langfuse's integration page documents the same 1.2.0 native export and says there is no wrapper to write. [1][2][3][4]
- The enable path is pip install "everos[otel]" plus an [observability] block. Tracing is off by default. Langfuse public/secret keys and host derive the OTLP endpoint and Basic auth; a raw endpoint can target any OTLP/HTTP backend. Content capture is opt-in. Environment equivalents are EVEROS_OBSERVABILITY__*. The chassis commit stamps langfuse.* attributes and continues an upstream W3C traceparent when present. [3][5][6][7][8]
- Each EverOS API call becomes one Langfuse trace. Writes map to everos.memory.add / everos.memory.flush, with child generations everos.memcell.boundary and everos.extract, persist span everos.persist.markdown, and later OME work under everos.ome.*, including everos.reflect.consolidate when related memories are merged and outdated ones retired. Searches map to retriever everos.memory.search over everos.search.recall and everos.search.rank, with everos.embedding nested in recall. Background reflection reattaches via traceparent so it stays inside the triggering write. [3][9][10]
- Recall quality is not on the OTLP stream. scores.py posts NUMERIC scores to {langfuse_host}/api/public/scores from a bounded background queue, only when Langfuse credentials and emit_recall_scores are set. Calibrated HYBRID and AGENTIC methods emit recall_top_score and recall_hit; KEYWORD and single-route VECTOR emit recall_top_score_raw. The default hit threshold is 0.6. A generic Collector will see search spans, not these scores. [11][5][12][13]
- Langfuse says LLM and embedding calls carry model and token usage so Langfuse computes cost; EverOS sends no cost of its own. The same page warns that recall-quality numbers are proxies, not ground-truth relevance, and recommends an LLM-as-judge over query and recalled content. 1.2.1 split raw versus calibrated names so Langfuse charts do not average unbounded BM25 with a [0, 1] probability. The example README records a real 1.2.1 demo as 237 spans over 60 traces. [3][13][9]
- Do not deploy 1.2.0 itself. The 1.2.1 release and CHANGELOG state that 1.2.0 omitted 1.1.4 fixes, including a knowledge-upload path traversal (CWE-22) that 1.1.4 had already closed. Native tracing survived; the security fix did not land until 1.2.1. Current PyPI is 1.5.1. Privacy default remains metadata-only unless capture_content is turned on. [13][2][4][5]

## What it means

If you already run EverOS beside a Langfuse-traced agent, this is the missing memory layer, not a new product to evaluate from scratch. Install everos[otel] at 1.2.1 or later (current 1.5.1), enable [observability], and propagate W3C traceparent from the agent so memory sits in the same trace as the LLM calls. Keep capture_content off unless you intend to send query and extracted-memory text. Chart recall_top_score only for HYBRID and AGENTIC; use recall_top_score_raw for KEYWORD and VECTOR; do not treat either as labeled relevance.

If you are not on Langfuse, you still get vendor-neutral OTLP spans and token usage, but not recall scores or a billed price. Langfuse's own OTLP endpoint is HTTP JSON or protobuf on /api/public/otel, not gRPC. For real quality, attach LLM-as-judge evals to the search traces rather than trusting recall_hit. Skip 1.2.0 on security grounds even though that is the version that introduced tracing.

## Still open

- No independent third-party audit of the traces was found; vendor docs and source agree, but production screenshots outside the cookbook were not verified here.
- Published CPU or latency overhead with tracing on was not located.
- How cleanly EverOS spans nest under Langfuse Python or JS SDK traces in mixed process topologies is not demonstrated beyond traceparent continuation.
- Whether other OTLP backends can reconstruct recall-hit dashboards without the Langfuse scores REST path remains open.

## Sources

1. [EverOS 1.2.0](https://github.com/EverMind-AI/EverOS/releases/tag/v1.2.0) (primary)
2. [EverOS CHANGELOG](https://raw.githubusercontent.com/EverMind-AI/EverOS/main/CHANGELOG.md) (primary)
3. [Observe the memory layer with EverOS and Langfuse](https://langfuse.com/integrations/other/everos) (primary)
4. [everos v1.5.1 on PyPI](https://pypi.org/project/everos/) (primary)
5. [EverOS default.toml observability settings](https://raw.githubusercontent.com/EverMind-AI/EverOS/main/src/everos/config/default.toml) (primary)
6. [feat(observability): OTel tracing chassis + request-id propagation](https://github.com/EverMind-AI/EverOS/commit/ab5cf0447ea5eff1338326dab1a12b3052bb77f7) (primary)
7. [everos v1.2.0 on PyPI](https://pypi.org/project/everos/1.2.0/) (primary)
8. [OpenTelemetry (OTEL) for LLM Observability](https://langfuse.com/integrations/native/opentelemetry) (primary)
9. [EverOS Langfuse native OpenTelemetry example README](https://raw.githubusercontent.com/EverMind-AI/EverOS/main/examples/langfuse/README.md) (primary)
10. [feat(observability): instrument memory ops + OME trace linking](https://github.com/EverMind-AI/EverOS/commit/357f619c64c5ce03e4c25b76f19e08f0528a0588) (primary)
11. [EverOS recall-score sink (Langfuse REST)](https://raw.githubusercontent.com/EverMind-AI/EverOS/main/src/everos/core/observability/tracing/scores.py) (primary)
12. [fix(observability): separate uncalibrated recall scores by name](https://github.com/EverMind-AI/EverOS/commit/4e13f7881e5bad0d20b31da62083fb6ff0b40eb0) (primary)
13. [EverOS 1.2.1](https://github.com/EverMind-AI/EverOS/releases/tag/v1.2.1) (primary)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Memory](https://newruntime.com/topics/memory.md): Explore the memory topic hub.
- topic hub: [Observability](https://newruntime.com/topics/observability.md): Explore the observability topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)